|Publication number||US7412428 B2|
|Application number||US 10/748,631|
|Publication date||Aug 12, 2008|
|Filing date||Dec 30, 2003|
|Priority date||Mar 12, 2002|
|Also published as||US20040162796|
|Publication number||10748631, 748631, US 7412428 B2, US 7412428B2, US-B2-7412428, US7412428 B2, US7412428B2|
|Original Assignee||Knowmtech, Llc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (77), Non-Patent Citations (101), Referenced by (21), Classifications (22), Legal Events (4)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This patent application is a continuation-in-part of U.S. patent application Ser. No. 10/095,273 filed on Mar. 12, 2002 now U.S. Pat. No. 6,889,216, the disclosure of which is incorporated herein by reference. This patent application is also a continuation-in-part of U.S. patent application Ser. No. 10/162,524 filed on Jun. 5, 2002 now abandoned, the disclosure of which is incorporated herein by reference. This patent application is additionally a continuation-in-part of U.S. patent application Ser. No. 10/226,191 filed on Aug. 22, 2002 now abandoned, the disclosure of which is incorporated herein by reference.
The present invention generally relates to molecular technology, including nanotechnology. The present invention also relates to neural networks and neural computing systems and teaching methods thereof. The present invention also relates to Hebbian and anti-Hebbian learning techniques applicable to neural networks.
Neural networks are computational systems that permit computers to essentially function in a manner analogous to that of the human brain. Neural networks do not utilize the traditional digital model of manipulating 0's and 1's. Instead, neural networks create connections between processing elements, which are equivalent to neurons of a human brain. Neural networks are thus based on various electronic circuits that are modeled on human nerve cells (i.e., neurons).
Generally, a neural network is an information-processing network, which is inspired by the manner in which a human brain performs a particular task or function of interest. Computational or artificial neural networks are thus inspired by biological neural systems. The elementary building blocks of biological neural systems are the neuron, the modifiable connections between the neurons, and the topology of the network.
Biologically inspired artificial neural networks have opened up new possibilities to apply computation to areas that were previously thought to be the exclusive domain of human intelligence. Neural networks learn and remember in ways that resemble human processes. Areas that show the greatest promise for neural networks, such as pattern classification tasks, speech and image recognition, are areas where conventional computers and data-processing systems have had the greatest difficulty.
In general, artificial neural networks are systems composed of many nonlinear computational elements operating in parallel and arranged in patterns reminiscent of biological neural nets. The computational elements, or nodes, are connected via variable weights that are typically adapted during use to improve performance. Thus, in solving a problem, neural net models can explore many competing hypothesis simultaneously using massively parallel nets composed of many computational elements connected by links with variable weights.
In contrast, with conventional von Neumann computers, an algorithm must first be developed manually, and a program of instructions written and executed sequentially. In some applications, this has proved extremely difficult. This makes conventional computers unsuitable for many real-time problems for which we have no efficient algorithm.
In a neural network, “neuron-like” nodes can output a signal based on the sum of their inputs, the output being the result of an activation function. In a neural network, there exists a plurality of connections, which are electrically coupled among a plurality of neurons. The connections serve as communication bridges among of a plurality of neurons coupled thereto. A network of such neuron-like nodes has the ability to process information in a variety of useful ways. By adjusting the connection values between neurons in a network, one can match certain inputs with desired outputs.
One does not program a neural network. Instead, one “teaches” a neural network by examples. Of course, there are many variations. For instance, some networks do not require examples and extract information directly from the input data. The two variations are thus called supervised and unsupervised learning. Neural networks are currently used in applications such as noise filtering, face and voice recognition and pattern recognition. Neural networks can thus be utilized as an advanced technique for processing information.
Neural networks that have been developed to date are largely software-based. A true neural network (e.g., the human brain) is massively parallel (and therefore very fast computationally) and very adaptable. For example, half of a human brain can suffer a lesion early in its development and not seriously affect its performance. Software simulations are slow because during the learning phase a standard computer must serially calculate connection strengths. When the networks get larger (and therefore more powerful and useful), the computational time becomes enormous.
For example, networks with 10,000 connections can easily overwhelm a computer. In comparison, the human brain has about 100 billion neurons, each of which can be connected to about 5,000 other neurons. On the other hand, if a network is trained to perform a specific task, perhaps taking many days or months to train, the final useful result can be built or “downloaded” onto a piece of hardware and also mass-produced. Because most problems requiring complex pattern recognition are highly specific, networks are task-specific. Thus, users usually provide their own, task-specific training data.
A number of software simulations of neural networks have been developed. Because software simulations are performed on conventional sequential computers, however, they do not take advantage of the inherent parallelism of neural network architectures. Consequently, they are relatively slow. One frequently used measurement of the speed of a neural network processor is the number of interconnections it can perform per second.
For example, the fastest software simulations available can perform up to approximately 18 million interconnects per second. Such speeds, however, currently require expensive super computers to achieve. Even so, approximately 18 million interconnects per second is still too slow to perform many classes of pattern classification tasks in real time. These include radar target classifications, sonar target classification, automatic speaker identification, automatic speech recognition, electrocardiogram analysis, etc.
The implementation of neural network systems has lagged somewhat behind their theoretical potential due to the difficulties in building neural network hardware. This is primarily because of the large numbers of neurons and weighted connections required. The emulation of even of the simplest biological nervous systems would require neurons and connections numbering in the millions and/or billions.
Due to the difficulties in constructing such highly interconnected processors, currently available neural network hardware systems have not approached this level of complexity. Another disadvantage of hardware systems is that they typically are often custom designed and configured to implement one particular neural network architecture and are not easily, if at all, reconfigurable in implementing different architectures. A true physical neural network chip, with the learning abilities and connectivity of a biological network, has not yet been designed and successfully implemented.
The problem with a pure hardware implementation of a neural network utilizing existing technology is the inability to physically form a great number of connections and neurons. On-chip learning can exist, but the size of the network is limited by digital processing methods and associated electronic circuitry. One of the difficulties in creating true physical neural networks lies in the highly complex manner in which a physical neural network must be designed and constructed. The present inventor believes that solutions to creating a true physical and artificial neural network lie in the use of nanotechnology and the implementation of a novel form of variable connections.
The term “Nanotechnology” generally refers to nanometer-scale manufacturing processes, materials and devices, as associated with, for example, nanometer-scale lithography and nanometer-scale information storage. Nanometer-scale components find utility in a wide variety of fields, particularly in the fabrication of microelectrical and microelectromechanical systems (commonly referred to as “MEMS”). Microelectrical nano-sized components include transistors, resistors, capacitors and other nano-integrated circuit components. MEMS devices include, for example, micro-sensors, micro-actuators, micro-instruments, micro-optics, and the like.
In general, nanotechnology presents a solution to the problems faced in the rapid pace of computer chip design in recent years. According to Moore's law, the number of switches that can be produced on a computer chip has doubled every 18 months. Chips now can hold millions of transistors. It is, becoming increasingly difficult, however, to increase the number of elements on a chip utilizing existing technologies. At the present rate, in the next few years the theoretical limit of silicon-based chips will likely have been attained. Because the number of elements and components that can be manufactured on a chip determines the data storage and processing capabilities of microchips, new technologies are required for the development of higher performance integrated circuit chips.
Present chip technology is also limited in cases where wires must be crossed on a chip. For the most part, the design of a computer chip is limited to two dimensions. Each time a circuit is forced to cross another circuit, another layer must be added to the chip. This increases the cost and decreases the speed of the resulting chip. A number of alternatives to standard silicon based complementary metal oxide semiconductor (“CMOS”) devices have been proposed. The common goal is to produce logic devices on a nanometer scale. Such dimensions are more commonly associated with molecules than integrated circuits.
The issue of interconnects in neural network hardware poses a serious problem. Because of the massive interconnectivity, a neural network constructed with standard integrated electronic methods can never reach the desired neuron and synapse density, simply because the interconnections overwhelm the largely 2-diminsional chip. It can thus be appreciated that almost any sort of 3-diminsional connectivity, no matter how simple, could offer tremendous benefits.
Integrated circuits and electrical components thereof, which can be produced at a molecular and nanometer scale, include devices such as carbon nanotubes and nanowires, which essentially are nanoscale conductors (“nanoconductors”). Nanoconductors are tiny conductive tubes (i.e., hollow) or wires (i.e., solid) with a very small size scale (e.g., 0.7 to 300 nanometers in diameter and up to 1 mm in length). Their structure and fabrication have been widely reported and are well known in the art. Carbon nanotubes, for example, exhibit a unique atomic arrangement, and possess useful physical properties such as one-dimensional electrical behavior, quantum conductance, and ballistic electron transport.
Carbon nanotubes are among the smallest dimensioned nanotube materials with a generally high aspect ratio and small diameter. High-quality single-walled carbon nanotubes can be grown as randomly oriented, needle-like or spaghetti-like tangled tubules. They can be grown by a number of fabrication methods, including chemical vapor deposition (CVD), laser ablation or electric arc growth.
Carbon nanotubes can be grown on a substrate by catalytic decomposition of hydrocarbon containing precursors such as ethylene, methane, or benzene. Nucleation layers, such as thin coatings of Ni, Co, or Fe are often intentionally added onto the substrate surface in order to nucleate a multiplicity of isolated nanotubes. Carbon nanotubes can also be nucleated and grown on a substrate without a metal nucleating layer by using a precursor including one or more of these metal atoms. Semiconductor nanowires can be grown on substrates by similar processes.
Attempts have been made to construct electronic devices utilizing nano-sized electrical devices and components. For example, a molecular wire crossbar memory device can be constructed from crossbar arrays of nanowires sandwiching molecules that act as on/off switches. Such a device can be formed from a plurality of nanometer-scale devices, each device comprising a junction formed by a pair of crossed wires where a single wire crosses another and at least one connector species connects the pair of crossed wires in the junction. The connector species comprises a bi-stable molecular switch. The junction forms either a resistor or a diode or an asymmetric non-linear resistor. The junction has a state that is capable of being altered by application of a first voltage and sensed by the application of a second, non-destructive voltage. A series of related patents attempts to cover everything from molecular logic to how to chemically assemble these devices.
Such a molecular crossbar device has two general applications. The notion of transistors built from nanotubes and relying on nanotube properties is being pursued. Second, two wires can be selectively brought to a certain voltage and the resulting electrostatic force attracts them. When they touch, the Van der Walls force keeps them in contact with each other and a “bit” is stored. The connections in this apparatus can therefore be utilized for a standard (i.e., binary and serial) computer. The inventors of such a device thus desire to coax a nanoconductor into a binary storage media or a transistor. As it turns out, such a device is easier to utilize as a storage device.
A need exists for a physical neural network, which can be implemented in the context of a semiconductor integrated circuit (i.e., a computer chip). Such a device, which can be referred to as a “physical neural network chip” or a “synapse chip” is thus disclosed herein. Such a device, if successfully implemented would be well suited for use with neural networks.
Researchers in the neuro-biological fields have attempted to develop a computationally efficient algorithm that can emulate a biologically realistic neural network. Specifically, researchers have attempted to develop method and/or systems, which would allow the efficient calculation of Spike-Timing Dependent-Plasticity (STDP), while also permitting fully interconnected pulsed networks. In STDP, timing between pre- and post-synaptic events can cause a net potentiation (LTP) or a net depression (LTD) of synapses.
A background for STPD is explained in “Mechanisms and Significance of Spike-Timing Dependent Plasticity” by Uma R. Karmarkar, et al., Biol. Cybern. 87, 373-382 (2002). STPD is also explained generally in “A Model of Spike-Timing Dependent Plasticity: One or Two Coincidence Detectors?” by Uma R. Karmarkar, et al., J. Neurophysiol, 88: 507-513, 2002. An additional reference regarding STDP is “Spike-Based Learning Rules and Stabilization of Persistent Neural Activity,” by Xiaohui Xie, et al., Dept. of Brain & Cog, Sci, MIT, Cambridge, Mass. A further reference regarding STDP and issues faced thereof is described in “Dendritic Spikes as a Mechanism for Cooperative Long-Term Potentiation,” Nace L. Golding, et al., Nature, Vol. 418, 18 Jul. 2002, pp. 326-330. The aforementioned references are indicated herein for generally illustrative, background and edification purposes only and do not disclose any features which limit any embodiments disclosed herein.
In the field of neural networks, Hebbian and anti-Hebbian learning has been the subject of great interest in recent years. An example of Hebbian learning is disclosed in non-limiting and non-essential U.S. Pat. No. 6,643,627, “Dynamic Synapse for Signal Processing in Neural Networks,” which issued to Liaw et al on Nov. 4, 2003, and which is incorporated herein by reference. Non-essential U.S. Pat. No. 6,643,627 is a non-essential reference with respect to embodiments of the present invention, and is referred to herein merely for general edification and background purposes only.
Another example of Hebbian learning is disclosed in non-limiting and non-essential U.S. Pat. No. 6,611,823, “Backlash Compensation Using Neural Network”, which issued to Selmic et al on Aug. 26, 2003. U.S. Pat. No. 6,111,823 also is a non-essential reference with respect to embodiments of the present invention, and is referred to herein merely for general edification and background purposes only. A further example of Hebbian learning and rules thereof is disclosed in non-essential U.S. Pat. No. 6,601,049, “Self-adjusting multi-layer neural network architectures and methods therefore”, which issued to David Cooper on Jul. 29, 2003, and is incorporated herein by reference. U.S. Pat. No. 6,601,049 is also a non-essential reference with respect to embodiments of the present invention, and is referred to herein merely for general edification and background purposes only.
To date, Hebbian and anti-Hebbian learning has been implemented utilizing software simulations and algorithms thereof. Some large scale physical neural network systems have also attempted to incorporate Hebbian and anti-Hebbian learning. Current Hebbian and anti-Hebbian learning systems have not been implemented via true physical neural networks, particularly networks developed at a molecular scale, such as physical neural networks implemented utilizing nanotechnology.
The following summary of the invention is provided to facilitate an understanding of some of the innovative features unique to the present invention, and is not intended to be a full description. A full appreciation of the various aspects of the invention can be gained by taking the entire specification, claims, drawings, and abstract as a whole.
It is, therefore, one aspect of the present invention to provide for a physical neural network, including an adaptive neural network, which can be formed and implemented utilizing nanotechnology.
It is still another aspect of the present invention to provide a physical neural network, which can be formed from a plurality of interconnected molecular connections, such as, for example, molecules, nanoconnections, and/or nanoconnectors.
It is yet a further aspect of the present invention to provide a physical neural network, which can be formed from a plurality of molecules, including molecular conducting structures.
It a further aspect of the present invention to provide a physical neural network based on nanoconductors, such as, for example, nanowires and/or nanotubes.
It is also an aspect of the present invention to provide a nanotechnology-based physical neural network, which incorporates Hebbian and anti-Hebbian learning techniques and mechanisms thereof.
The above and other aspects can be achieved as will now be described. Methods and systems are disclosed herein in which a physical neural network can be configured utilizing nanotechnology. Such a physical neural network can comprise a plurality of molecular conductors (e.g., nanoconductors) which form neural connections between pre-synaptic and post-synaptic components of the physical neural network. Additionally, a learning mechanism can be applied for implementing Hebbian learning via the physical neural network. Such a learning mechanism can utilize a voltage gradient or voltage gradient dependencies to implement Hebbian and/or anti-Hebbian plasticity within the physical neural network. Such a learning mechanism can also utilize pre-synaptic and post-synaptic frequencies to provide Hebbian and/or anti-Hebbian learning within the physical neural network. Examples of nanoconductors which can be utilized to implement nanoconnections of the physical neural network include nanoconductors such as nanotubes, nanowires, nanoparticles and the like.
In general, A system can be implemented which comprises a plurality of molecular conductors disposed within a dielectric medium, at least one input electrode in contact with the dielectric medium, at least one output electrode in contact with the dielectric medium, wherein the plurality of molecular conductors form physical neural connections when the dielectric medium is exposed an electric field across the at least one input electrode and the at least one output electrode, such that the physical neural connections can be strengthened or weakened depending upon a strengthening or weakening of the electric field or an alteration of a frequency thereof. The physical neural network can therefore include plurality of molecular conductors disposed within the dielectric medium, the input electrode(s) in contact with the dielectric medium, and the output electrode(s) in contact with the dielectric medium.
The accompanying figures, in which like reference numerals refer to identical or functionally-similar elements throughout the separate views and which are incorporated in and form part of the specification, further illustrate the present invention and, together with the detailed description of the invention, serve to explain the principles of the present invention.
The particular values and configurations discussed in these non-limiting examples can be varied and are cited merely to illustrate an embodiment of the present invention and are not intended to limit the scope of the invention.
The physical neural network described and disclosed herein is different from prior art forms of neural networks in that the disclosed physical neural network does not require computer calculations for training, nor is its architecture based on any current neural network hardware device. The physical neural network described herein is generally fast and adaptable, no matter how large such a physical neural network becomes.
The physical neural network described herein with respect to one or more embodiments can be referred to generically as a Knowm™. The terms “physical neural network” and “Knowm” can thus be utilized interchangeably to refer to the same device, network, or structure. The term “Knowm” can also refer to a semiconductor implementation, such as a physical neural network chip and/or synapse chip. Note that the terms “physical neural network chip” and “synapse chip” can also be utilized herein to refer generally to the same or analogous type of Knowm™ device.
Network orders of magnitude larger than current VSLI neural networks can now be built. One consideration for a Knowm™ is that it must be large enough for its inherent parallelism to shine through. Because the connection strengths of such a physical neural network are dependant on the physical movement of nanoconnections thereof, the rate at which a small network can learn is generally very small and a comparable network simulation on a standard computer can be very fast. On the other hand, as the size of the network increases, the time to train the device does not change. Thus, even if the network takes a full second to change a connection value a small amount, if it does the same to a billion connections simultaneously, then its parallel nature begins to express itself.
A physical neural network (i.e., a Knowm™ enabled physical neural network) generally should possess two basic components to function properly. First, such a physical neural network should have one or more neuron-like nodes that sum a signal and output a signal based on the amount of input signal received. Such a neuron-like node is generally non-linear in output. In other words, there should be a certain threshold for input signals, below which nothing can be output and above which a constant or nearly constant output is generated or allowed to pass. This is considered the basic building block of all neural networks, and can be accomplished by an activation function. The second requirement of a physical neural network is the inclusion of a connection network composed of a plurality of interconnected electrodes (i.e., nanoconnections). Such a connection network is described in greater detail herein.
As illustrated in
In a Knowm™, the neuron-like node can be configured as a standard diode-based circuit, the diode being the most basic semiconductor electrical component, and the signal it sums can be a voltage. An example of such an arrangement of circuitry is illustrated in
Although a diode may not necessarily be utilized, its current versus voltage characteristics are non-linear when used with associated resistors and similar to the relationship depicted in
Thus, neuron 200 comprises a neuron-like node that may include a diode 206, which is labeled D1, and a resistor 204, which is labeled R2. Resistor 204 is connected to a ground 210 and an input 205 of diode 206. Additionally, a resistor 202, which is represented as a block and labeled R1, can be connected to input 205 of diode 206. Block 202 includes an input 212, which comprises an input to neuron 200. A resistor 208, which is labeled R3, is also connected to an output 214 of diode 206. Additionally, resistor 208 is coupled to ground 210. Diode 206 in a physical neural network is analogous to a neuron of a human brain, while an associated connection formed thereof, as explained in greater detail herein, is analogous to a synapse of a human brain.
As depicted in
An amplifier may also replace diode 206 so that the output voltage immediately saturates at a reference threshold voltage, thus resembling a step function. R3 (i.e., resistor 208) functions generally as a bias for diode 206 (i.e., D1) and should generally be about 10 times larger than resistor 204 (i.e., R2). In the circuit configuration illustrated in
Nanoconnections 304, which are analogous to biological synapses, can be composed of electrical conducting material (i.e., nanoconductors). Nanoconductors can be provided in a variety of shapes and sizes without departing from the teachings herein. A nanoconductor can also be implemented as, for example, a molecule or groups of molecules. A nanoconductor can also be implemented as, for example, DNA. Studies have shown that DNA has special electrical properties which can function as essentially a tiny electrical wire. This recent discovery opens up a possible route to new applications in the electronics industry and particularly with respect to the physical neural network disclosed herein.
Carbon particles (e.g., granules or bearings) can also be utilized for developing nanoconnections. The nanoconductors utilized to form a connection network can be formed as a plurality of nanoparticles. For example, each nanoconnection within a connection network can be formed from a chain of carbon nanoparticles. Thus, nanoconductors that are utilized to form a physical neural network (i.e., Knowm™) can be formed from such nanoparticles. Note that as utilized herein, the term “nanoparticle” can be utilized interchangeably with the term “nanoconductor.” The term “nanoparticle” can refer simply to a particular type of nanoconductors, such as, for example, a carbon nanoparticle, or another type of nanoconductors, such as, for example, a carbon nanotube or carbon nanowire. Devices that conduct electricity and have dimensions on the order of nanometers can be referred to as nanoconductors.
A connection network as disclosed herein can be composed from a variety of different types of nanoconductors. For example, a connection network can be formed from a plurality of nanoconductors, including nanowires, nanotubes and/or other types of nanoparticles or molecular conductors. Note that such nanowires, nanotubes and/or nanoparticles, along with other types of nanoconductors can be formed from materials such as carbon or silicon. For example, carbon nanotubes may comprise a type of nanotube that can be utilized in accordance with the present invention.
As illustrated in
A connection network is thus not one connection between two electrodes, but a plurality of connections between input electrodes and output electrodes. Nanotubes, nanowires, nanoparticles and/or other nanoconducting structures can be utilized, of course, to construct nanoconnections 304 between input 302 and input 306. Although a single input 302 and a single input 306 is depicted in
The connection network also comprises a plurality of interconnected nanoconnections, wherein each nanoconnection thereof is strengthened or weakened according to an application of an electric field. A connection network is not possible if built in one layer because the presence of one connection can alter the electric field so that other connections between adjacent electrodes could not be formed. Instead, such a connection network can be built in layers, so that each connection thereof can be formed without being influenced by field disturbances resulting from other connections. This can be seen in
Nanoconnections 572, 574, 576, 578, and 580 may comprise nanoconductors such as, for example, nanotubes and/or nanowires. Nanoconnections 572, 574, 576, 578, and 580 thus comprise one or more nanoconductors. Additionally, input lines 522, 524, 526, 528, and 530 are respectively coupled to a plurality of input lines 542, 544, 546, 548 and 550, which are in turn each respectively coupled to nanoconnections 582, 584, 586, 588, and 590.
Thus, for example, input line 542 is connected to nanoconnections 582, while input line 544 is connected to nanoconnections 584. Similarly, input line 546 is connected to nanoconnections 586 and input line 548 is connected to nanoconnections 588. Additionally, input line 550 is connected to nanconnections 590. Box 556 and 554 generally represent simply the output and are thus illustrated connected to outputs 562 and 568. In other words, outputs 556 and 554 respectively comprise outputs 562 and 568. The aforementioned input lines and associated components thereof actually comprise physical electronic components, including conducting input and output lines and physical nanoconnections, such as nanotubes and/or nanowires.
Thus, the number of layers 558 and 560 equals the number of desired outputs 562 and 568 from physical neural network 500. In the previous two figures, every input was potentially connected to every output, but many other configurations are possible. The connection network can be made of any electrically conducting material, although the practicality of the application requires that they be very small so that they will align with a practical voltage. Carbon nanotubes or any conductive nanowire can be implemented in accordance with the physical neural network described herein. Such components can thus form connections between electrodes by the presence of an electric field.
The only general requirements for the conducting material utilized to configure the nanoconductors are that such conducting material must conduct electricity, and a dipole should preferably be induced in the material when in the presence of an electric field. Alternatively, the nanoconductors utilized in association with the physical neural network described herein can be configured to include a permanent dipole that is produced by a chemical means, rather than a dipole that is induced by an electric field. A connection network could also be configured from other conductive particles that are developed or found useful in the nanotechnology arts. For example, carbon particles (e.g., carbon “dust”) may also be used as nanoconductors in place of nanowires or nanotubes. Such particles may include bearings or granule-like particles.
A connection network can be constructed as follows. Initially, a voltage can be applied across a gap that is filled with a mixture of nanowires and a “solvent”. This mixture can be composed of a variety of materials or substances. The only general requirement in constructing such a connection network is that the conducting wires should be suspended in the solvent and/or dissolved or in a suspension, but free to move about. Additionally, the electrical conductance of the substance should generally be less than the electrical conductance of the suspended conducting nanowire(s) and/or other nanoparticle(s). The viscosity of the substance should not be too much so that the conducting nanowire(s) and/or other nanoparticles(s) cannot move when an electric field is applied.
The goal for such a connection network is to develop a network of connections of just the “right” values so as to satisfy particular signal-processing requirements, which is precisely how a neural network functions. Applying a voltage across a space occupied by the aforementioned mixture can form a connection network. To create a connection network, input terminals can be selectively raised to a positive voltage, while the output terminals can be selectively grounded.
Alternatively, an electric field, either AC or DC, can be applied across the terminals. Such an electric field can be, for example, a sinusoidal, square or a saw-tooth waveform. Thus, connections can gradually form between the inputs and outputs. The important requirement that makes a physical neural network functional as a neural network in accordance with an embodiment of the present invention is that the longer this electric field is applied across the connection gap, and/or the greater the frequency or amplitude of the field, the more nanotubes and/or nanowires and/or nanoparticles align and the stronger the connections thereof become.
The connections can either be initially formed and possess random resistances or no connections may be formed at all. By initially forming random connections, it might be possible to teach the desired relationships faster, because it is not necessary for the base connections to be constructed from scratch. Depending on the rate of connection decay, having initial random connections could prove faster, although not necessarily. The connection network can adapt itself to the requirements of a given situation regardless of the initial state of the connections.
The resistance of a connection can be maintained or lowered by selective activations of the connection. In other words, an electric field can be applied perpendicular to the direction of connection formation by perpendicular electrodes. Alternately, both the input and output electrodes could be given the same sinusoidal, alternating signal, which would create pulses of electrostatic repulsion in the connection region. The temperature of the solution can also be controlled so that the rate that connection degradation can be controlled.
The nanoconnections may or may not be arranged in an orderly array pattern between the input and output electrodes. The nanoconnections (e.g., nanotubes, nanowires, etc) of a physical neural network do not have to order themselves into neatly formed arrays. They simply float in the solution, or lie at the bottom of the gap, and more or less line up in the presence an electric field. Precise patterns are thus not necessary. In fact, neat and precise patterns may not be desired. Rather, precise patterns could be a drawback rather than an advantage. In fact, it may be desirable that the connections themselves function as poor conductors, so that variable connections are formed thereof, overcoming simply an “on” and “off” structure, which is commonly associated with binary and serial networks and structures thereof.
Although it can be seen that nanoparticles aligned in a dielectric medium (e.g., a dielectric solvent or dielectric solution) can offer a unique solution to emulated modifiable, variable connections within an electronic implementation of a neural network, it is not yet obvious how one would provide feedback that would train the connections. A training mechanism may be implemented in many different forms. Basically, the connections in a connection network must be able to change in accordance with the feedback provided. In other words, the very general notion of connections being strengthened or connections being weakened in a physical system is the essence of a physical neural network (i.e., a Knowm™ physical neural network).
Thus, it can be appreciated that the training of such a physical neural network may not require a “CPU” to calculate connection values thereof. The Knowm™ physical neural network, including artificial synapses thereof, can adapt itself. Complicated neural network solutions could be implemented very rapidly “on the fly”, much like a human brain adapts as it performs. It is anticipated that various learning mechanisms can be implemented in accordance with preferred or alternative embodiments of the present invention. Two such learning mechanisms are generally discussed herein. First, a feedback mechanism is described that leads to the training of a multi-layer, feed-forward network. Second, a feedback mechanism is generally discussed, which can result in Hebbian synapse modification within recurrent, highly interconnected networks.
The physical neural network disclosed herein thus has a number of broad applications. The core concept of a Knowm™ physical neural network, however, is basic. The very basic idea that the connection values between electrode junctions by nanoconductors can be used in a neural network device is all that required to develop an enormous number of possible configurations and applications thereof.
An important feature of a physical neural network is the ability to form negative connections. This is an important feature that makes possible inhibitory effects useful in data processing. The basic idea is that the presence of one input can inhibit the effect of another input. In artificial neural networks as they currently exist, this is accomplished by multiplying the input by a negative connection value. Unfortunately, with a physical device, the connection may only take on zero or positive values under such a scenario
In other words, either there can be a connection or no connection. A connection can simulate a negative connection by dedicating a particular connection to be negative, but one connection cannot begin positive and through a learning process change to a negative connection. In general, if it starts positive, it can only go to zero. In essence, it is the idea of possessing a negative connection initially that results in the simulation, because this does not occur in a biological network. Only one type of signal travels through axons and dendrites in a biological network. That signal is transferred into the flow of a neurotransmitter whose effect on the receiving neuron can be either excitatory or inhibitory, depending on the neuron, thereby dedicating certain connections inhibitory and excitatory
One method for solving this problem is to utilize two sets of connections for the same output, having one set represent the positive connections and the other set represent the negative connections. The output of these two layers can be compared, and the layer with the greater output will output either a high signal or a low signal, depending on the type of connection set (inhibitory or excitatory). Such a configuration be seen in
A truth table for the output of circuit 700 is illustrated at block 780 in
At all times during the learning process, a weak alternating electric field can be applied perpendicular to the connections. This can cause the connections to weaken by rotating the nanotube perpendicular to the connection direction. This weakening of connections is important because it can allow for a much higher degree of adaptation. To understand this, one must realize that the connections cannot (practically) keep getting stronger and stronger. By weakening those connections not contributing much to the desired output, we decrease the necessary strength of the needed connections and allow for more flexibility in continuous training. Other mechanisms, such as increasing the temperature of the nanotube suspension could also be used for such a purpose.
The circuit depicted in
The second part of the circuit (i.e., the connections) is thus a new type of chip structure, although it could be constructed with current technology. The connection chip can be composed of an orderly array of electrodes spaced anywhere from, for example, 100 nm to 1 μm or perhaps even further. In a biological system, one talks of synapses connecting neurons. It is in the synapses where the information is processed, (i.e., the “connection weights”). Similarly, such a chip can contain all of the synapses for the physical neural network. A possible 2-diminsional arrangement thereof can be seen in
The training of such a chip is primarily based on two assumptions. First, the inherent parallelism of a physical neural network (i.e., a Knowm™ network or system) can permit all training sessions to occur simultaneously, no matter how large the associated connection network. Second, recent research has indicated that near perfect aligning of nanotubes can be accomplished in no more than 15 minutes utilizing practical voltages of about 5V.
If one considers that the input data, arranged as a vector of binary “high's” and “low's” is presented to the Knowm™ network or system simultaneously, and that all training vectors are presented one after the other in rapid succession (e.g., perhaps 100 MHz or more), then each connection would “see” a different frequency in direct proportion to the amount of time that its connection is required for accurate data processing (i.e., provided by a feedback mechanism). Thus, if it only takes approximately 15 minutes to attain an almost perfect state of alignment, then this amount of time would comprise the longest amount of time required to train, assuming that all of the training vectors are presented during that particular time period and adequate feedback has been provided.
The solvent utilized can comprise a volatile liquid that can be confined or sealed and not exposed to air. For example, the solvent and the nanoconductors present within the resulting solution can be sandwiched between wafers of silicon or other materials. If the fluid has a melting point that is approximately at operating temperature, then the viscosity of the fluid could be controlled easily. Thus, if it is desired to lock the connection values into a particular state, the associated physical neural network (i.e., a Knowm™ network or system) can be cooled slightly until the fluid freezes. The term “solvent” as utilized herein thus can include fluids such as for example, toluene, hexadecane, mineral oil, liquid crystals, etc. Note that the solution in which the nanoconductors (i.e., nanoconnections) are present should generally comprise a substance that does not conduct electricity and allows for the suspension of nanoparticles.
Thus, when the resistance between the electrodes is measured, the conductivity of the nanoconductors can be measured, not that of the solvent. The nanoconductors can be suspended in the solution or can alternately lie on the bottom surface of the connection gap if the gap is 2-diminsional (i.e., formed on a planar surface such as electrodes deposited on the surface of a substrate). Note that the solvent and/or solutions described herein may also comprise liquid crystal media.
As illustrated thereafter at block 906, the nanoconductors must be suspended in the solvent, either dissolved or in a suspension of sorts, but generally free to move around, either in the solution or on the bottom surface of the gap. As depicted next at block 908, the electrical conductance of the solution must be less than the electrical conductance of the suspended nanoconductor(s). Next, as illustrated at block 910, the viscosity of the substance should not be too much so that the nanoconductors cannot move when an electric field (e.g., voltage) is applied across the electrodes. Finally, as depicted at block 912, the resulting solution of the “solvent” and the nanoconductors is thus located within the connection gap.
Note that although a logical series of steps is illustrated in
As illustrated thereafter at block 1006, connections thus form between the inputs and the outputs. The important requirements that make the resulting physical neural network functional as a neural network is that the longer this electric field is applied across the connection gap, or the greater the frequency or amplitude, the more nanoconductors align and the stronger the connection becomes. Thus, the connections that experience the most feedback during training become the strongest.
As indicated at block 1008, the connections can either be initially formed and have random resistances or no connections will be formed at all. By forming initial random connections, it might be possible to teach the desired relationships faster, because the base connections do not have to be built up as much. Depending on the rate of connection decay, having initial random connections could prove to be a faster method, although not necessarily. A connection network will adapt itself to whatever is required regardless of the initial state of the connections.
Thus, as indicated at block 1010, as the electric field is applied across the connection gap, the more the nonconductor(s) will align and the stronger the connection becomes. Connections (i.e., synapses) that are not used are dissolved back into the solution, as illustrated at block 1012. As illustrated at block 1014, the resistance of the connection can be maintained or lowered by selective activations of the connections. In other words, “if you do not use the connection, it will fade away,” much like the connections between neurons in a human brain in response to Long Term Depression, or LTD.
The neurons in a human brain, although seemingly simple when viewed individually, interact in a complicated network that computes with both space and time. The most basic picture of a neuron, which is usually implemented in technology, is a summing device that adds up a signal. Actually, this statement can be made even more general by stating that a neuron adds up a signal in discrete units of time. In other words, every group of signals incident upon the neuron can be viewed as occurring in one moment in time. Summation thus occurs in a spatial manner. The only difference between one signal and another signal depends on where such signals originate. Unfortunately, this type of data processing excludes a large range of dynamic, varying situations that cannot necessarily be broken up into discrete units of time.
The example of speech recognition is a case in point. Speech occurs in the time domain. A word is understood as the temporal pronunciation of various phonemes. A sentence is composed of the temporal separation of varying words. Thoughts are composed of the temporal separation of varying sentences. Thus, for an individual to understand a spoken language at all, a phoneme, word, sentence or thought must exert some type of influence on another phoneme, word, sentence or thought. The most natural way that one sentence can exert any influence on another sentence, in the light of neural networks, is by a form of temporal summation. That is, a neuron “remembers” the signals it received in the past.
The human brain can accomplish such a feat in an almost trivial manner. When a signal reaches a neuron, the neuron has an influx of ions rush through its membrane. The influx of ions contributes to an overall increase in the electrical potential of the neuron. Activation is achieved when the potential inside the cell reaches a certain threshold. The one caveat is that it takes time for the cell to pump out the ions, something that it does at a more or less constant rate. So, if another signal arrives before the neuron has time to pump out all of the ions, the second signal will add with the remnants of the first signal and achieve a raised potential greater than that which could have occurred with only the second signal. The first signal influences the second signal, which results in temporal summation.
Implementing this in a technological manner has proved difficult in the past. Any simulation would have to include a “memory” for the neuron. In a digital representation, this requires data to be stored for every neuron, and this memory would have to be accessed continually. In a computer simulation, one must discritize the incoming data, since operations (such as summations and learning) occur serially. That is, a computer can only do one thing at a time. Transformations of a signal from the time domain into the spatial domain require that time be broken up into discrete lengths, something that is not necessarily possible with real-time analog signals in which no point exists within a time-varying signal that is uninfluenced by another point.
A physical neural network, however, is generally not digital. A physical neural network is a massively parallel analog device. The fact that actual molecules (e.g., nanoconductors) must move around (in time) makes one form of temporal summation a natural occurrence. This temporal summation is built into the nanoconnections and can occur at a time scale much longer that that which is possible with capacitors and standard analog circuitry in micron dimension VLSI designs. The easiest way to understand this is to view the multiplicity of nanoconnections as one connection with one input into a neuron-like node (Op-amp, Comparator, etc.). This can be seen in
Input 1102 can be provided by another physical neural network (i.e., a Knowm™ network or system) to cause increased connection strength of nanoconnections 1104 over time. This input will most likely arrive in pulses, but can also be continuous, depending upon a desired implementation. A constant or pulsed electric field perpendicular to the connections would serve to constantly erode the connections, so that only signals of a desired length or amplitude could cause a connection to form.
Once the connection is formed, the voltage divider formed by nanoconnection 1104 and resistor 1106 can cause a voltage at node A in direct proportion to the strength of nanoconnections 1104. When the voltage at node A reaches a desired threshold, the amplifier (i.e., an op-amp and/or comparator), will output a high voltage (i.e., output 1114). The key to the temporal summation is that, just like a real neuron, it takes time for the electric field to breakdown the nanoconnections 1104, so that signals arriving close in time will contribute to the firing of the neuron (i.e., op-amp, comparator, and forth). Temporal summation has thus been achieved. The parameters of the temporal summation could be adjusted by the amplitude and frequency of the input signals and the perpendicular electric field.
The CPU 1211 can perform various processing and controlling functions, such as pattern recognition, including but not limited to speech and/or visual recognition based on the output signals from the physical neural network device 1222. The CPU 1211 is connected to a read-only memory (ROM) 1213, a random-access memory (RAM) 1214, a communication control unit 1215, a printer 1216, a display unit 1217, a keyboard 1218, an FFT (fast Fourier transform) unit 1221, a physical neural network device 1222 and a graphic reading unit 1224 through a bus line 1220 such as a data bus line. The bus line 1220 may comprise, for example, an ISA, EISA, or PCI bus.
The ROM 1213 is a read-only memory storing various programs or data used by the CPU 1211 for performing processing or controlling the learning process, and speech recognition of the physical neural network device 1222. The ROM 1213 may store programs for carrying out the learning process according to error back-propagation for the physical neural network device or code rows concerning, for example, 80 kinds of phonemes for performing speech recognition. The code rows concerning the phonemes can be utilized as second instructor signals and for recognizing phonemes from output signals of the neuron device network. Also, the ROM 1213 can store programs of a transformation system for recognizing speech from recognized phonemes and transforming the recognized speech into a writing (i.e., written form) represented by characters.
A predetermined program stored in the ROM 1213 can be downloaded and stored in the RAM 1214. RAM 1214 generally functions as a random access memory used as a working memory of the CPU 1211. In the RAM 1214, a vector row storing area can be provided for temporarily storing a power obtained at each point in time for each frequency of the speech signal analyzed by the FFT unit 1221. A value of the power for each frequency serves as a vector row input to a first input portion of the physical neural network device 1222. Further, in the case where characters or graphics are recognized in the physical neural network device, the image data read by the graphic reading unit 1224 are stored in the RAM 1214.
The communication control unit 1215 transmits and/or receives various data such as recognized speech data to and/or from another communication control unit through a communication network 1202 such as a telephone line network, an ISDN line, a LAN, or a personal computer communication network. Network 1202 may also comprise, for example, a telecommunications network, such as a wireless communications network. Communication hardware methods and systems thereof are well known in the art.
The printer 1216 can be provided with a laser printer, a bubble-type printer, a dot matrix printer, or the like, and prints contents of input data or the recognized speech. The display unit 1217 includes an image display portion such as a CRT display or a liquid crystal display, and a display control portion. The display unit 1217 can display the contents of the input data or the recognized speech as well as a direction of an operation required for speech recognition utilizing a graphical user interface (GUI).
The keyboard 1218 generally functions as an input unit for varying operating parameters or inputting setting conditions of the FFT unit 1221, or for inputting sentences. The keyboard 1218 is generally provided with a ten-key numeric pad for inputting numerical figures, character keys for inputting characters, and function keys for performing various functions. A mouse 1219 can be connected to the keyboard 1218 and serves as a pointing device.
A speech input unit 1223, such as a microphone can be connected to the FFT unit 1221. The FFT unit 1221 transforms analog speech data input from the voice input unit 1223 into digital data and carries out spectral analysis of the digital data by discrete Fourier transformation. By performing a spectral analysis using the FFT unit 1221, the vector row based on the powers of the respective frequencies are output at predetermined intervals of time. The FFT unit 1221 performs an analysis of time-series vector rows, which represent characteristics of the inputted speech. The vector rows output by the FFT 1221 are stored in the vector row storing area in the RAM 1214.
The graphic reading unit 224, provided with devices such as a CCD (Charged Coupled Device), can be used for reading images such as characters or graphics recorded on paper or the like. The image data read by the image-reading unit 1224 are stored in the RAM 1214. Note that an example of a pattern recognition apparatus, which can be modified for use with the physical neural network of the present invention, is disclosed in non-essential U.S. Pat. No. 6,026,358 to Tomabechi, Feb. 16, 2000, “Neural Network, A Method of Learning of a Neural Network and Phoneme Recognition Apparatus Utilizing a Neural Network,” which is incorporated herein by reference. U.S. Pat. No. 6,026,358 is referenced herein only for edification and illustrative purposes only.
The implications of a physical neural network are tremendous. With existing lithography technology, many electrodes in an array such as depicted in
For example, such a chip can be constructed utilizing a standard computer processor in parallel with a large physical neural network or group of physical neural networks. A program can then be written such that the standard computer teaches the neural network to read, or create an association between words, which is precisely the same sort of task in which neural networks can be implemented. This would amount to nothing more than presenting the Knowm network with a pre-defined sequence of input and output patterns stored in memory.
Once the physical neural network is able to “read”, it can be taught for example to “surf” the Internet and find material of any particular nature. A search engine can then be developed that does not search the Internet by “keywords”, but instead by meaning. This idea of an intelligent search engine has already been proposed for standard neural networks, but until now has been impractical because the network required was too big for a standard computer to simulate. The use of a physical neural network as disclosed herein now makes a truly intelligent search engine possible.
A physical neural network can be utilized in other applications, such as, for example, speech recognition and synthesis, visual and image identification, management of distributed systems, self-driving cars and filtering. Such applications have to some extent already been accomplished with standard neural networks, but are generally limited in expense, practicality and not very adaptable once implemented.
The use of a physical neural network can permit such applications to become more powerful and adaptable. Indeed, anything that requires a bit more “intelligence” could incorporate a physical neural network. One of the primary advantages of a physical neural network is that such a device and applications thereof can be very inexpensive to manufacture, even with present technology. The lithographic techniques required for fabricating the electrodes and channels there between has already been perfected and implemented in industry.
Most problems in which a neural network solution is generally implemented are complex adaptive problems, which change in time. An example is weather prediction. The usefulness of a physical neural network is that it could handle the enormous network needed for such computations and adapt itself in real-time. An example wherein a physical neural network (i.e., Knowm™) can be particularly useful is the Personal Digital Assistant (PDA). PDA's are well known in the art. A physical neural network applied to a PDA device can be advantageous because the physical neural network can ideally function with a large network that could constantly adapt itself to the individual user without devouring too much computational time from the PDA processor. A physical neural network could also be implemented in many industrial applications, such as developing a real-time systems control to the manufacture of various components. This systems control can be adaptable and totally tailored to the particular application, as necessarily it must.
The training of multiple connection networks between neuron layers within a multi-layer neural network is an important feature of any neural network. The addition of neuron layers to a neural network can increase the ability of the network to create increasingly complex associations between inputs and outputs. Unfortunately, the addition of extra neuron layers in a network raises an important question: How does one optimize the connections within the hidden layers to produce the desired output?
The neural network field was stalled for some time trying to answer this question until several parties simultaneously stumbled onto a computationally efficient solution, now referred to generally as “back-propagation” or “back-prop” for short. As the name implies, the solution involves a propagation of error back from the output to the input. Essentially, back-propagation amounts to efficiently determining the minimum of an error surface composed of n variables, where the variable n represents the number of connections.
Because back propagation is a computational algorithm, this concept may not make much sense physically. Another related question to ask is can the neurons in a human brain take a derivative? Do they “know” the result of a connection on another neuron? In other words, how does a neuron know what the desired output is if each neuron is an independent summing machine, only concerned with its own activation level and firing only when that activation is above threshold? What exactly can a neuron “know” about its environment?
Although this question is certainly open for debate, it is plausible to state that a neuron can only “know” if it has fired and whether or not its own connections have caused the firing of other neurons. This is precisely the Hebbian hypothesis for learning: “if neuron A repeatedly takes part in firing neuron B, then the connection between neuron A and B strengthens so that neuron A can more efficiently take part in firing neuron B”. With this hypothesis, a technique can be derived to train a multi-layer physical neural network device without utilizing back-propagation or any other training “algorithm”, although the technique mirrors back-propagation in form, as information is transferred from output layers to input layers, providing feedback in the form of pulses that modify connections.
In fact, the resulting Knowm™ (i.e., physical neural network) can be self-adaptable and does not require any calculations. In other words, the network and its training mechanism can be a physical process that arises from feedback signals within the network. The structure of a Knowm™ physical neural network or synapse thereof thus creates a situation in which learning simply takes place when a desired output is given. The description that follows is thus based on the use of a physical neural network (i.e., a Knowm™) and constituent nanoconnections thereof.
Inputs 1304 and 1302 (i.e. I1 and I2) generally provide one or more signals, which can be propagated through connection network 1302 (i.e., C1). Connection network 1302 thus generates a first output signal at node 1303 and a second output signal at node 1305. The first output signal provided at node 1303 is further coupled to an input 1323 of an amplifier 1312, while the signal output signal provided at node 1305 is connected to an input 1325 of an amplifier 1314. Amplifier 1312 thus includes two inputs 1323 and 1311, while amplifier 1314 includes two inputs 1315 and 1325. Note that a voltage Vt can be measured at input 1311 to amplifier 1312. Similarly, voltage Vt can also be measured at input 1315 to amplifier 1314. Additionally, a resistor 1316 can be coupled to node 1305 and a resistor 1310 is connected to node 1303. Resistor 1310 is further coupled to a ground 1309. Resistor 1316 is further connected to ground 1309. Resistors 1310 and 1316 are labeled Rb in
Amplifier 1312 can thus function as a neuron A and amplifier 1314 functions as a neuron B. The two neurons, A and B, respectively sum the signals provided at nodes 1303 and 1305 to provide output signals thereof at nodes 1319 and 1321 (i.e., respectively H1 and H2). Additionally, a switch 1308, which is labeled S1, is connected between nodes 1303 and 1319. Likewise, a switch 1322, which is also labeled S1, is connected between nodes 1305 and 1321. A resistor 1318 is coupled between an output of amplifier 1312 and node 1319. Similarly, a resistor 1320 is coupled between an output of amplifier 1314 and node 1321. Node 1319, which carries signal H1, is connected to a connection network 1328. Also, node 1321, which carries signal H2, is connected to connection network 1328.
Note that connection network 1328 is labeled C2 in
A voltage Vt can be measured at an input 1335 to amplifier 1336 and an input 1337 to amplifier 1338. Amplifiers 1336 and 1338 can be respectively referred to as neurons C and D. An output from amplifier 1336 is connected to a NOT gate 1340, which provides a signal that is input to a NOR gate 1342. Additionally, amplifier 1338 provides a signal, which can be input to NOR gate 1342. Such a signal, which is output from amplifier 1338 can form an inhibitory signal, which is input to NOR gate 1342. Similarly, the output from amplifier 1336 can comprise an excitatory signal, which is generally input to NOT gate 1340. The excitatory and inhibitory signals respectively output from amplifiers 1336 and 1338 form an excitatory/inhibitory signal pair. NOR gate 1342 generates an output, which is input to an amplifier 1344 at input node 1347. A voltage Vd can be measured at input node 1346, which is coupled to amplifier 1344.
Thus, the signals H1 and H2, which are respectively carried at nodes 1319 and 1321, are generally propagated through connection network 1328, which is labeled C2, where the signals are again summed by the two neurons, C and D (i.e., amplifiers 1336 and 1338). The output of these two neurons therefore form an excitatory/inhibitory signal pair, which through the NOT gate 1340 and the NOR gate 1342 are transformed into a signal output O1 as indicated at output 1348. Note that signal output node O1 can be measured at input node 1347 of amplifier 1344. Amplifier 1344 also includes an output node 1349, which is coupled to node 1331 through a switch 1350, which is labeled S2. Output 1349 is further coupled to a NOT gate 1354, which in turn provides an output which is coupled to node 133 through a switch 1352, which is also labeled S2.
For inhibitory effects to occur, it may be necessary to implement twice as many outputs from the final connection network as actual outputs. Thus, every actual output represents a competition between a dedicated excitatory signal and inhibitory signal. The resistors labeled Rb (i.e., resistors 1330 and 1334) are generally very large, about 10 or 20 times as large as a nanoconnection. On the other hand, the resistors labeled Rf (i.e., resistors 1318 and 1320) may possess resistance values that are generally less than that of a nanoconnection, although such resistances can be altered to affect the overall behavior of the associated physical neural network. Vt represents the threshold voltage of the neuron while Vd represents the desired output. S1 and S2 are switches involved in the training of layers 1 and 2 respectively (i.e., L1 and L2, which are indicated respectively by brackets 1326 and 1356 in
For reasons that will become clear later, a typical training cycle can be described as follows. Initially, an input vector can be presented at I1 and I2. For this particular example, such an input vector generally corresponds to only 4 possible combinations, 11, 10, 01 or 00. Actual applications would obviously require many more inputs, perhaps several thousand or more. One should be aware that the input vector does not have to occur in discrete time intervals, but can occur in real time. The inputs also need not necessarily be digital, but for the sake of simplicity in explaining this example, digital representations are helpful. While an input pattern is being presented, a corresponding output can be presented at Vd. Again, in this particular case there is only one output with only two corresponding possible outcomes, 1 or 0. The desired output also does not have to be presented in discrete units of time.
For learning to occur, the switches 1350 and 1352 (i.e., S2) can be closed, followed by switches 1308 and 1322 (i.e., S1). Both groupings of switches (S1 and S2) can then be opened and the cycle thereof repeated. Although only two layers L1 and L2 are illustrated in
For example, it can be assumed that no connections have formed within connection networks C1 or C2 and that inputs are being matched by desired outputs while the training wave is present. Since no connections are present, the voltage at neurons A, B, C and D are all zero and consequently all neurons output zero. One can quickly realize that whether the training wave is present or not, a voltage drop will not ensue across any connections other than those associated with the input connection network. The inputs, however, are being activated. Thus, each input is seeing a different frequency. Connections then form in connection network C1, with the value of the connections essentially being random.
Before a connection has been made, the voltage incident on neurons A and B is zero, but after a connection has formed, the voltage jumps to approximately two diode drops short of the input voltage. This is because the connections form a voltage divider with Rb, such that Rb (i.e., resistors 1310 and/or 1316) possesses a resistance very much larger than that of the nanoconnections. Two reasons for utilizing a large Rb are to minimize power consumption of the physical neural network during a normal operation thereof, and to lower the voltage drop across the connections so that learning (i.e. connection modification) only takes place when feedback is present. Fortunately, nanotube contact resistances are on the order of about 100 kΩ, or more, which can allow for an Rb of a few MΩ or greater. Vt must be somewhere between two diode drops of the input voltage and the voltage produce by one nanoconnection in a voltage divider with Rb, the later being lower than the former.
Once connections have formed across C1 and grown sufficiently strong enough to activate neurons A and B, the connections across C2 can form in the same manner. Before continuing, however, it is important to determine what will occur to the nanoconnections of connection network 1302 (i.e., C1) after they grow strong enough to activate the first layer neurons. For the sake of example, assume that neuron A has been activated. When S1 is closed in the training wave, neuron A “sees” a feedback that is positive (i.e., activated). This locks the neuron into a state of activation, while S1 is closed. Because of the presence of diodes in connection network 1302 (i.e., C1), current can only flow from left to right in C1. This results in the lack of a voltage drop across the nanoconnections.
If another electric field is applied at this time to weaken the nanoconnections (e.g., perhaps a perpendicular field), the nanoconnections causing activation to the neuron can be weakened (i.e., the connections running from positive inputs to the neuron are weakened) This can also be accomplished by an increased temperature, which could naturally arise from heat dissipation of the other circuitry on the chip. This feedback will continue as long as the connections are strong enough to activate the neuron (i.e., and no connections have formed in the second layer). Nanoconnections can thus form and be maintained at or near the values of neuron activation. This process will also occur for ensuing layers until an actual network output is achieved.
Although the following explanation for the training of the newly formed (and random) connections may appear unusual with respect to
It can be appreciated from
Recall, however, that in a typical physical neural network many more outputs are likely to be utilized. If the output is high when the desired output is low, then the training neuron (i.e., amplifier 1344, the last neuron on the right in
Because of the presence of diodes within connection network 1328 (i.e., C2), there will be no voltage drop across those connections going to the excitatory neuron. There will be a voltage drop, however, across the nanoconnections extending from positive inputs of C2 to the inhibitory neuron (i.e., amplifier 1338). This can result in increases in inhibitory nanoconnections and a decrease in excitatory nanoconnections thereof (i.e., if an eroding is present). This is exactly what is desired if the desired output is low when the actual output is high. A correspondingly opposite mechanism strengthens excitatory connections and weakens inhibitory connections if the desired output is high when the actual output is low. When the desired output matches the actual output, the training neurons output is dependent on the gain of the differential amplifier.
Thus far an explanation has been presented describing how the last layer of a physical neural network can in essence train itself to match the desired output. An important concept to realize, however, is that the activations coming from the previous layer are basically random. Thus, the last connection network tries to match essentially random activations with desired outputs. For reasons previously explained, the activations emanating from the previous layer do not remain the same, but fluctuate. There must then be some way to “tell” the layers preceding the output layer which particular outputs are required so that their activations are no longer random.
One must realize that neurons simply cannot fire unless a neuron in a preceding layer has fired. The activation of output neurons can be seen as being aided by the activations of neurons in previous layers. An output neuron “doesn't care” what neuron in the previous layer is activating it, so long as it is able to produce the desired output. If an output neuron must produce a high output, then there must be at least one neuron in the previous layer that both has a connection to it and is also activated, with the nanoconnection(s) being strong enough to allow for activation, either by itself or in combination with other activated neurons.
With this in mind, one can appreciate that the nanoconnections associated with pre-output layers can be modified. Again, by referring to
By thereafter closing S1, the previous layer neurons in essence “know” how much of their activation signal is being utilized. If their signal is being utilized by many neurons in a preceding layer, or by only a few with very strong nanoconnections, then the voltage that the neuron receives as feedback when S1 is closed decreases to a point below the threshold of the neuron. Exactly what point this occurs at is dependent on the value of Rf (i.e., resistors 1318 and/or 1320) As Rf becomes larger, less resistance is generally required to lower H1 or H2 to a point below the threshold of the neuron. This feedback voltage is very important, as this is how the network matches inputs with desired outputs. First, note that the feedback is local, confined to individual neurons. In essence, if a neuron needs to supply activations to many neurons, then it must strengthen its connections to neurons that are activating it, so that it may continue to do its job.
Several subtleties exist in this feedback process. Although the feedback voltage is largely determined by the neurons' pre-synaptic connections, (i.e., “axonal” connections), it is also determined by the neurons' post synaptic connections (i.e., dendritic connections). If the feedback voltage, Vf, is lower than the threshold voltage, Vt, then the dendritic connections will be strengthened. Because the feedback voltage is a function of both the axonal and dendritic connections, one scenario that cannot lower Vf below Vt is weak axonal connections and very strong dendritic connections. In other words, if the dendritic connections (to activated neurons) are very strong, then the axonal connections (to neurons needing activation) must be correspondingly stronger. This relationship is not linear. Thus, based on the foregoing, nanoconnections in layers preceding the output layer can modify themselves.
Referring again to
Note that connections could also form between activated pre-synaptic neurons and the neuron in question even if no initial connection is present, or if the post-synaptic neuron is inactivated. This last form of connection formation is important because it allows for a form of connection exploration. In other words, connections can be formed, and if the feedback mechanism finds it useful to match a desired input-output relationship, it will be strengthen. If not, it will be weakened. This allows neuron A to keep outputting a high signal that in turn allows the output neurons to match the desired output. The same argument can apply for neuron B, or any neuron in any layer preceding the output layer.
Although a detailed description of the process has been provided above, it is helpful to view the process from a generalized perspective. Again, assuming that no connections are present in any of the connection networks, assume that a series of input vectors are presented to the inputs of the network, and a series of output vectors are presented to the desired output, while the training wave is present. The training wave should be at a frequency equal or greater than the frequency at which input patterns are presented or otherwise the first few layers will not be trained and the network will be unable to learn the associations. The first layer connection network, analogous to C1 in
The connections can, just like C1, build up and hover around the threshold voltage for the succeeding neurons. This pattern of forming connections can generally occur until a signal is achieved at the output. Once a signal has been outputted, the feedback process begins and the training wave guides the feedback so that connections are modified strategically, from the output connection network to the input connection network, to achieve the desired output. The training is continued until the user is satisfied with the networks ability to correctly generate the correct output for a given input.
In evaluating a standard feed-forward multi-layer neural network, connections generally form between every neuron in one layer and every neuron in the next layer. Thus, neurons in adjacent layers are generally completely interconnected. When implementing this in a physical structure where connection strengths are stored as a physical connection, the architecture must be configured that allows for both total connectedness between layers and which also provides for the efficient use of space. In a physical neural network device (i.e., a “Knowm™ device), connections form between two conducting electrodes. The space between the electrodes can be filled with a nano-conductor/dielectric solvent mixture, which has been described previously herein. As an electric field is applied across the electrode gap, connections form between the electrodes. A basic method and structure for generating a large number of synapses on a small area substrate is illustrated in
The basic structure of a physical neural network device, such as a physical neural network chip and/or synapse chip, is depicted in
The input electrodes are indicated in
Although only four input electrodes and four output electrodes are illustrated in
A larger view of an adaptive network system can thus be seen in
The base neurons can be in a perpendicular array structure (i.e., chip 1758) composed of two or more layers 1752, 1754 coupled with Knowm synapses (i.e., system 1700). Each Knowm synapse can be composed of connection conduits, separated by a characteristic distance “d”, where each connection conduit is the result of nano-particles aligning in an electric field generated by the temporal and sequential firing of the coupled base neurons (i.e., see schematic diagram 1756). External inputs to the network can be coupled to any post-synaptic electrode of any base neuron in any layer. And any network output can be provided at any pre-synaptic electrode of any base neuron in any layer.
Other attempts at creating a neural-like processor require components to be placed precisely, with resolutions of a nanometer. The design of
Although the electrode dimensions cannot be lowered indefinitely without a considerable loss in connection resistance variation, it is conceivable that a 1 cm2 chip could hold over 4 billion synapses (e.g., 50 nm electrodes and 100 nm spacing=4.4 billion synapes/cm2). Because neuron circuitry could potentially be constructed on the other side of the synapse arrays, very compact neural processors with high neuron/synapse density could also be constructed.
Some considerations about the construction of a chip should be addressed. For example, the distance between the input electrodes should generally remain at a distance close, but not touching, the output electrodes. If carbon nanotubes are utilized for the nano-conductors within the gaps, one would need to prepare the nanotubes to lengths shorter than the gap distance. If the gap distance is, for example, approximately, 100 nm, then the nanotubes should be sized less this dimension. Given a diameter of about 1.5 nm, nanotubes can only go so far, perhaps 10's of nanometers. At 1.5 nm, one is now approaching atomic distances. The distance between the two electrodes may be maintained by resting the upper plate of electrodes on “pedestals”, which can be formed by an interference photolithography technique
Note that as utilized herein, the term “chip” can generally refer to a type of integrated circuit device comprising a large number of connected circuit elements such as transistors and resistors, fabricated on a single chip of silicon crystal or other semiconductor material. Such chips have traditionally been manufactured as flat rectangular or square shaped objects. It can be appreciated, however, that such chips can be fabricated in a variety of shapes, including circular and spherical shapes in addition to traditional square, box or rectangular shaped integrated circuit chips. Thus, a synapse chip or physical neural network chip (i.e., a Knowm™ chip) can also be fabricated as a spherical integrated circuit.
An example of a spherical chip is disclosed in non-limiting and non-essential U.S. Pat. No. 6,245,630, “Spherical Shaped Semiconductor Circuit,” which issued to Akira Ishikawa of Ball Semiconductor, Inc. on Jun. 12, 2001 and which is incorporated herein by reference The spherical chip disclosed in U.S. Pat. No. 6,245,630 generally comprises a spherical shaped semiconductor integrated circuit (“ball”) and a system and method for manufacturing the same. Thus, a ball can replace the function of a flat, conventional chip. Note that U.S. Pat. No. 6,245,630 is disclosed herein for general edification, illustrative and background purposes only and is not considered a limiting reference with respect to the embodiments disclosed herein.
The physical dimensions of the ball allow it to adapt to many different manufacturing processes which otherwise could not be used. Furthermore, the assembly and mounting of the ball may facilitate efficient use of the semiconductor as well as circuit board space. Thus, a physical neural network chip and/or synapse chip as disclosed herein can be configured as such a ball-type chip in addition to a rectangular or square shaped integrated circuit chip.
Based on the foregoing it can be appreciated that embodiments are generally directed toward a physical neural network synapse chip and also a method for forming such a synapse chip. The synapse chip disclosed herein with respect to one or more embodiments can be configured to include an input layer comprising a plurality of input electrodes and an output layer comprising a plurality of output electrodes, such that the output electrodes are located above or below the input electrodes. A gap is generally formed between the input layer and the output layer. A solution can then be provided which is prepared from a plurality of nanoconductors and a dielectric medium (e.g., a dielectric solvent or solution).
The solution can be located within the gap, such that an electric field is applied across the gap from the input layer to the output layer to form nanoconnections of a physical neural network implemented by the synapse chip. Such a gap can thus be configured as an electrode gap. The input electrodes can be configured as an array of input electrodes, while the output electrodes can be configured as an array of output electrodes.
The nanoconductors can form nanoconnections at one or more intersections between the input electrodes and the output electrodes in accordance with an increase in strength of the electric field applied across the gap from the input layer to the output layer. Additionally, an insulating layer can be associated with the input layer, and another insulating layer associated with the output layer. The input layer can be formed from a plurality of parallel N-type semiconductors and the output layer formed from a plurality of parallel P-type semiconductors.
Similarly, the input layer can be formed from a plurality of parallel P-type semiconductors and the output layer formed from a plurality of parallel N-type semiconductors. Thus, the nanoconnections can be strengthened or weakened respectively according to an increase or a decrease in strength of the electric field. As an electric field is applied across the electrode gap, nanoconnections thus form between the electrodes.
The most important aspect of the electrode arrays described herein is their geometry. Generally, any pattern of electrodes in which almost every input electrode is connected to every output electrode, separated by a small gap, is a valid base for a connection network. What makes this particular arrangement better than other arrangements is that it is very space-efficient. By allowing the connection to form vertically, a third dimension can be utilized, consequently gaining enormous benefits in synapse density.
To understand just how space-efficient a Knowm™ chip utilizing connection formation in a third-dimension could be, consider the NET talk network created by Terry Sejnowski and Charles Rosenberg in the mid 1980's. NET talk took the text-representation of a word and could output the phonemic representation, thereby providing a text-to-speech translation. Such a network provided 203 inputs, 120 hidden neurons and 26 outputs, for a total of approximately 28 thousand synapses.
Utilizing electrode widths of approximately 200 nm and spacing between electrodes of approximately 400 nm, one could contain 28 thousand synapses on about 10160 μm2. In comparison, a conventional synapse including all of the weight storage resistors and switches, 10-14 current mirrors, multiplier and sign switching circuitry can take up, for example, up approximately 106◊113 μm or 11978 μm2. In other words, one could fit 28 thousand synapses in less than the area previously required to store only one synapse.
The benefits of creating a neural network processor are thus great. The ability to implement as many as 1 billion synapses on 1 cm2 of surface substrate is a tremendous leap forward over prior art neural network technologies. Another innovation is the ability to mass-produce pre-trained, large-scale neural network chips. A physical neural network as disclosed herein does not have to be taught at all, but can instead be manufactured with the desired connections already in place. This is an important feature for consumer devices. For example, in most cellular telephones produced today, the ability to recognize rudimentary speech is available. One might, after pre-recording a voice, speak the word “Dave” and the cellular telephone can automatically call Dave after matching the word just spoken to a list of other pre-recorded names and thereafter pick the best match.
This is a rather rudimentary form of pattern recognition and could therefore be replaced by an exceedingly small Knowm™ synapse chip. For example, a Knowm™ chip can be taught at the factory to translate speech into text, thereby eliminating the need to pre-record ones voice for recognition tasks and instead relying on a more general speech recognition technique. Once the factory Knowm™ chip is trained, the synapse resistance values can be determined. With knowledge of what each synapse value needs to be, one can then design a perpendicular array chip so that the electrode widths create a cross-sectional area inversely proportional to the resistance of each synapse.
In other words, the resistance of each connection is generally a function of the cross-sectional area of the connection between electrodes. By pre-forming the electrodes to certain specified widths, and then allowing the maximum number of connections to form at each electrode intersection, a physical neural network can be mass-produced. Such a configuration can allow a very general network function (e.g., voice or facial recognition) to be produced and sold to consumers, without the necessity of forcing the consumer to train the network.
A synapse or physical neural network chip could therefore be produced with certain ready-made abilities, such as voice or facial recognition. After installation, it is up to the designer to create a product that can then modify itself further and continue to adapt to the consumer. This could undoubtedly be an advantageous ability. Utilizing the example of the cellular telephone, the cellular telephone could in essence adapt its speech-recognition to the accent or manner of speech of the individual user. And all of this is possible because the Knowm™ synapses are so space-efficient. Networks with very powerful pattern recognition abilities could fit into a tiny fraction of a hand-held device, such as, for example, a wireless personal digital assistant and/or a cellular telephone.
The embodiments indicated herein are thus directed toward a physical neural network that can be configured from a connection or a plurality of connections of molecules or molecular conductors, such as nanoconductors, such as, for example, nanowires, nanotubes, and/or nanoparticles. Such physical neural network can be implemented in the form of one or more synapse chips that can be combined with a neuron system (e.g., a neuron chip) of independent summing circuits.
The fundamental concept of a Knowm™ network or system (e.g., a Knowm™ synapse chip) is remarkably simple. When particles in a dielectric solution are exposed to an electric field (i.e., AC or DC), the particles can align with the field. As the particles align, the resistance between the respective electrodes decreases. The connection becomes stable once the electric field is removed. As the strength or frequency of the applied electric field is increased, the connections become increasingly aligned and the resistance further decreases. By applying a perpendicular electric field, one can also decrease the strength of the connections. Such connections can be utilized as “synapses” in a physical neural network chip (also referred to as a synapse chip), and the result is a Knowm™ enabled neural network chip—a fully adaptable, high-density neural network chip.
Although a multi-layer, feed-forward structure, has been addressed, recurrent, unsupervised neural networks can provide many advantages. An adaptive neuron, as a base for a larger Hebbian-based recurrent network utilizing Knowm™ synapses will be discussed. The basic theory behind adaptive networks will be presented; after which will be described the appropriate translation to an electrical system that utilizes Knowm™ synapses.
A Knowm™ synapse can be configured in a manner that is highly appropriate for an adaptive neural network, which can also be referred to as an adaptive integration network or simply, an adaptive network. As indicated earlier, adaptive neural networks to date have been limited to software designs and/or conventional hardware implementations. Adaptive neural networks have not been designed or implemented based on nanotechnology components, systems, and/or networks as discussed herein.
Such neurons can be implemented as a Knowm™ network or system. Each neuron 1910, 1920, 1930, 1940, 1950, and 1960 can be individually formed from standard photolithography or alternate procedures by building circuits capable of neuronal function, as will be later discussed. Alternatively, connections between neurons 1910, 1920, 1930, 1940, 1950, and 1960 may be formed as nanoconductor(s) suspended within a dielectric solvent or solution. An example of nanoconnections that may be implemented to neurons 1910, 1920, 1930, 1940, 1950, and 1960 is provided by nanoconnections 304 of
A neuron combined with pre-synaptic electrodes can thus be the basic processing element of an adaptive integration network and can be (although not necessarily) configured to receive signals from its “pre-synaptic” neurons as input and, in response, transmit output signals to its “post-synaptic” neurons. A neuron has two output states, firing or non-firing. In one embodiment, binary output signal values of xi=1 and xi=0 are assigned for the firing and non-firing states, some embodiments may employ non-binary values for the output signal xi, for example, within a range 0.0≦x1≦1.0.
As another example, the value of the output signal xi can be 0.0 if the neuron does not “fire”, and greater than or equal to 1.0 if the neuron does fire. In the context of electronic circuitry, the output of the neuron can comprise a voltage. When a neuron fires, that neuron could potentially cause its post-synaptic neurons to fire, as more specifically explained herein after, which could cause their post-synaptic neurons to fire, and so on, setting up a chain reaction along an active pathway.
Any neuron in an adaptive integration network can be designated as a data input neuron or a data output neuron. A data input neuron is a neuron that receives a signal external to the adaptive integration network, and a data output neuron is a neuron whose output signal is transmitted to a destination external to the adaptive integration network. Accordingly, external signals input into data input neurons may initiate a chain reaction of neuron firings throughout the adaptive integration network. When the neuron firings eventually affect the state of the data output neurons, the output of the adaptive integration network will change in response.
In the example of
In an adaptive neural network, a connection is the conduit along which a neuron receives a signal from another neuron. Connections can be formed between neurons in any direction to transmit a signal from an output of a pre-synaptic neuron to an input of a post-synaptic neuron. Typically, a neuron plays both roles, first as a post-synaptic neuron for receiving input signals from its pre-synaptic neurons, and second as a pre-synaptic neuron for generating output signals to its post-synaptic neurons.
For example, with continued reference to
Connections (e.g., nanoconnections) may be excitatory or inhibitory, through which transmitted signals respectively promote or retard the firing of the post-synaptic neuron in response. With continued reference to
Excitatory connections are used to transmit signals from one neuron to another in a feedback loop or other active pathway. Inhibitory connections, on the other hand, prevent neurons from firing and are useful in providing internal regulation among feedback loops, but cannot actually form a connection in a feedback loop. In the context of a physical neural network, the inhibitory connections may cause a momentary increase in the threshold voltage of the post-synaptic neuron, thereby inhibiting the activations of the neuron.
An adaptive integration network may be configured to include feedback loops. A loop is a closed circuit of linked excitatory connections arranged in the same circular direction. For example, adaptive integration network 1900 comprises two loops, a first loop with neurons 1910, 1940, and 1950 indicated with black excitatory connections 1914, 1945, and 1951, and a second loop with neurons 1920, 1930, and 1960 denoted with gray excitatory connections 1923,1936, and 1962.
Loops are highly interactive with other loops. In general, a loop can be mutually reinforcing or mutually competitive with one or more other loops. The adaptive integration network 1900 depicted in
In addition, neurons 1910 and 1950 have inhibitory connections 1912 and 1956, respectively for suppressing firings of neurons 1920 and 1960, respectively, in the second (gray) loop. Thus, activation of the first (black) loop can force the deactivation of the second (gray) loop. Similarly, activating the second (gray) loop builds a circular chain of firings through neurons 1920, 1930, and 1960, while suppressing activity in neurons 1910 and 1940, via inhibitory connections 1961 and 1934, respectively.
Mutually interacting loops may be aggregated to form metaloops at a higher level of integration. For example, two mutually interacting loops may share one or more connections in common such that activity in one loop will affect the activity in the other loop. Referring to
Loop 2002 comprises six neurons 2010, 2020, 2030, 2040, 2050, and 2060 connected in sequence, and loop 2004 comprises five neurons 250, 2060, 2070, 2080, and 2090 connected in sequence. Both loop 2002 and 2004 share neurons 2050 and 2060, which are coupled by connection 2056. Activity on either loop influences activity on the other loop. For example, if neuron 2010 in loop 2002 fires, that firing eventually results in the firing of neuron 2060, which transmits a signal to neuron 2010 of loop 2002 and to neuron 2070 of loop 2004. Similarly, if neuron 2070 in loop 2004 fires, that firing eventually results in the firing of neuron 2060, which transmits a signal to neuron 2010 of loop 2002 and to neuron 2070 of loop 2004.
As another example, one loop could branch off an active pathway to another loop, thereby initiating activity in the other loop.
Since an adaptive integration network provides much flexibility in configuration, it is to be understood that the present invention is not limited to any particular configuration of neurons and connections. Preferably, it is desirable to choose the number, distribution, and types of connections to maximize the total number of feedback loops while minimizing the functional constraints and interdependence of the loops. In general, this goal can be met by employing as many connections per node as feasible for a given implementation.
The distribution of connections can vary from implementation to implementation of an adaptive integration network. For example, a maximum length can limit connections so that distant neurons are not directly connected, and the assignment of connections can be determined randomly or in accordance with an algorithm designed to give each neuron a similar physical or logical arrangement.
In an adaptive integration network, a neuron fires in response to firings of the neuron's pre-synaptic neurons under certain conditions. More specifically, each neuron has an associated excitation level
Furthermore, each connection can be characterized by a corresponding synaptic efficiency in transferring its signal, represented by a connection weight wi, where i indicates the ith connection for the neuron. In the context of a physical neural network, the synaptic efficiency is a direct result of the alignment of the nano-conductors between pre-synaptic and post-synaptic electrodes, and the alignment is in turn a result of heightened activation.
In a hardware implementation, when a pre-synaptic neuron fires a signal to its post-synaptic neurons, the firing neuron causes the excitation level
In one embodiment, neurons may be subject to a refractory period in which the neuron's excitation level
In a hardware implementation, the following sub-circuits that compose an individual neuron accomplish the refractory period. The inputs from synaptic activations are summed via an integrator, which allows the accumulation of signals over time. The integrated signal is passed to a threshold circuit, such as a comparator or operational amplifier that outputs a high or low voltage in response to the integrator signal being above a set threshold. This signal is passed to a circuit that allows a pulse of period “T” to be generated. The output pulse is the output of the neuron. This output is feed into a refractory pulse generator, which serves the purpose of grounding the post-synaptic electrodes in a Knowm synapse while the neuron is actively generating a pulse.
If the output pulse of the neuron was high, then the refractory pulse generator could comprise a NOT gate, for example. The grounding of the postsynaptic electrodes serves two purposed. First, the neuron is re-set to a zero level activation, as described earlier. Second, the lowered potential causes an increase in the electric field across all connection in a connection network currently activating the neuron. In other words, during the time of the refractory pulse, all the connections that are coming from firing neurons become stronger.
Training is generally the process of updating the nano-connections in an adaptive integration network so that the adaptive integration network produces desired outputs in response to inputs. In contrast with prior techniques involving artificial neural networks that employ distinct training and implementation phases, training the adaptive integration network is constantly occurring during the normal operation of the adaptive integration network and is a direct result of feedback within the network.
Prior to operation of the adaptive integration network, the connection weights within the adaptive integration network are initialized, for example, either randomly or to a preset value. During the operation of the adaptive integration network, the connection weights are constantly strengthened or weakened, provided that the connection weight strengthening or weakening conditions are met. Connection weight strengthening refers to the process of decreasing the resistance of the nano-connection. Connection weight strengthening occurs whenever any two connected neurons fire in close temporal proximity, with the post-synaptic neuron firing after the pre-synaptic neuron, during the post-synaptic neurons refractory pulse period.
Optionally in some adaptive network implementations, connection weight strengthening occurs every time a neuron fires, but the magnitude of the connection weight strengthening is a function of the amount of time since the pre-synaptic neuron of the connection has fired. This is a natural result of frequency dependence on connection formation in a Knowm synapse since connections contributing less to the over-all activation of a neuron will receive fewer “refractory” pulses and consequently see a decreased frequency of electric field across the pre and post-synaptic electrode terminals.
Connection weight strengthening allows for frequently used neuronal pathways to be reinforced. As one neuron fires, the neuron produces an output signal that may induce one or more of the neuron's post-synaptic neurons to fire in close temporal proximity, thereby strengthening the connection between the neurons. Similarly, the firing of the post-synaptic neuron may cause that neuron's post-synaptic neuron to fire, creating a chain reaction of firing neurons along an active pathway. Since a connection weight increases when both the pre-synaptic and the post-synaptic neurons fire in close temporal proximity, each time the active neural pathway is used, the connection weights along the active pathway are increased.
A loop is a special case of a frequently used active pathway, because, once initiated, the neurons in the loop successively fire in cycles around the loop. Each time the neurons fire, their connections are strengthened, yielding a stable loop circuit. Consequently, the connection weight strengthening rules foster stable circuits of self-reinforcing loops, which can constitute stored memory of patterns and other information
Connection weight weakening generally refers to the process of decreasing the strength of the connection. In an adaptive network, connection weight weakening occurs after a specified period of passivity for the connection. A connection is considered “passive” for particular point in time if the post-synaptic neuron and the pre-synaptic neuron of the connection have not fired in close temporal proximity in that period of time. Thus, the connection weights for passive connections progressively weaken, reducing the influence of those passive connections in the adaptive integration network. In a physical Knowm implementation, a decrease in synapse activations results in a lower frequency of applied electric field and thus a decrease in connection formation. As discuss previously, The connection formation could be constantly degraded by a perpendicular electric field or even from a dissolution process within the solution.
Larger connection weights are slowly decreased, thereby allowing for strong connections to remain more or less fixed, slow to decay even after prolonged passivity. This effect is naturally achieved in a knowm network by a decrease in the local electric filed around a strong nano-connection, thereby weakening effects from perpendicular electric fields. Alternatively, a strong nano-connection results in higher van-der-wall attractions and a corresponding heightened resistance to dissolution within the dielectric medium. In an adaptive integration network, connection weights are constantly being adjusted during normal operation, for example, strengthened when two connected neurons fire in close temporal proximity or weakened after a period of passivity. Therefore, even mere use of the adaptive integration network causes the adaptive integration network to be fine-tuned.
In certain cases, however, it is desirable to cause the adaptive integration network to learn and adapt to new patterns and information.
As indicated at decision block 2302, a loop can be controlled long as the output data does not match the desired output. The network activity of the adaptive integration network can be increased, as depicted at block 2304, which causes the output data to change. When the desired data is produced, the network activity is restored to a normal level, as described at block 2306. Various techniques may be employed to increase network activity, i. e. the rate of neural firings, including threshold lowering and neural output signal magnification.
Network activity can be increased by lowering the thresholds of the neurons in the adaptive integration network. For example, the thresholds could be reduced by a fixed amount or proportion, such as to one half. Threshold lowering causes neurons to fire sooner, because the excitation level
Yet another technique for increasing network activity is to increase the magnitude of the neural signals. Each time a neuron fires, the excitation level
Increasing network activity enables for new active pathways to be explored. For example, a neuron that is adjacent to an active pathway, but not part of the active pathway, might not ordinarily fire because it has low connection strength for a connection to a neuron on the active pathway. In this case, the excitation level
Increasing network activity can also cause an active pathway for one stable circuit to transform into an active pathway for another stable circuit. A stable circuit, which constitutes stored memory, information, or patterns within the adaptive integration network, represents a locally optimal position in the solution space (all possible outputs for all possible input). As a result, increasing network activity permits adaptive exploration through the solution space in search of other locally optimal positions for the new input/output data sets. Another result of increasing network activity is that the response time of the adaptive integration network is reduced, making the adaptive integration network faster.
When network activity is increased, for example by lowering the threshold, the accumulated excitation level
Neuron 2450, moreover, is the source of an inhibitory connection to neuron 2420 of the active pathway, which prevents neuron 2420 from firing so often. As both neuron 2450 and neuron 2420 fire, the inhibitory connection between the two neurons is strengthened, further preventing neuron 2420 from firing so often. Eventually, the passivity of neuron 2420 causes the connection between neuron 2410 and 2420 to weaken, completing the dislodging of the active pathway.
These new pathways and loops will eventually affect the data output neurons and alter the output values. If the output values are still undesired, then the increase in the network activity is maintained, causing the new pathways and loops to be ephemeral and generating even newer active pathways and loops. As soon as the desired output is attained, the user can discontinue the network activity increasing signal, causing the relevant network parameters (thresholds, etc.) to rebound to their ordinary levels and ceasing the adaptive training. This process can be automated if the desired output is presented before hand so that the output of the adaptive integration network can be compared by computer with the desired output to generate the external signal.
In contrast with retraining methods for conventional artificial neural networks, including both software and hardware implementations thereof, adaptive learning with adaptive integration networks is less disruptive, particularly when implemented via nanotechnology devices and techniques, such as discussed herein. For example, with conventional artificial neural networks every neuron is perturbed during training, but with adaptive integration networks only the neurons along active pathways and their neighbors are affected.
Thus, only relevant connections are adjusted, and previously established but unrelated loops and meta-loops are left intact, which hold previously learned classifications and information. Therefore, in further contrast with conventional artificial neural networks, a nanotechnology-based adaptive integration network can learn new information and patterns at any time without having to relearn previously learned material or going through a new training stage.
Such signals may be generated in the form of a voltage or a current, depending upon a desired implementation. For illustrative purposes only, it can be assumed that such signals comprise voltage signals. As indicated next at block 2606, these signals provided by the connection network can be summed by at least one neuron within the physical neural network and then, as illustrated at block 2608, compared to a threshold value. The threshold voltage can be an externally applied and modifiable voltage.
If, as indicated at block 2610, the current state of activation does not exceed the threshold value or threshold voltage, the process simply terminates, as indicated at block 2611. If, however, the current state of activation does exceed the threshold value or threshold voltage, then the process continues, as indicated at block 2612, and a pulse (e.g., a voltage pulse or current pulse) is emitted from a neuron within the physical neural network. During this pulse, a “refractory pulse generator” grounds the postsynaptic junction thereof.
This operation in turn can cause the synapses receiving pre-synaptic activation to experience an increase in the local electric field. The pre-synaptic electrodes of succeeding neurons and post-synaptic connections of the pulse emitting neuron thus receive a pulse, as indicated at block 2615. Thus, synapses that contribute to the activation of the neuron can receive an increase in the local electric field parallel to the connection direction and can also experience a higher frequency of activation, two parameters that increase the strength nanoconnections thereof, as indicated at block 2618, and thus the strength of the Knowm™ synapse.
For illustrative purposes only, it can be assumed that the first neuron is fired. The firing of the first neuron causes an increase in the voltage of a pre-synaptic connection (e.g., a pre-synaptic electrode), as indicated at block 2708, and the an activation of a subsequent or second neuron, as illustrated at bock 2708, which in turn causes a refractory pulse to decrease the voltage of the post-synaptic connection (e.g., a post-synaptic electrode), as illustrated at block 2710. These operations in turn can generally result in an increased voltage between pre-synaptic electrodes and post-synaptic electrodes thereof, as depicted at block 2712. As indicated at block 2712, the processes illustrated beginning at block 2704 can be repeated for subsequent electrodes.
The result of the operations described at blocks 2704 to 2712 occurring many times in succession can produce, as illustrated at block 2716, an increased frequency of the electric field between the pre-synaptic and post-synaptic electrodes, thereby causing, as depicted at block 2718, an increase in the alignment of nanoparticles (e.g., nanotubes, nanowires, etc.) and a decrease in the electrical resistance between electrodes thereof. The process can then terminate, as indicated at block 2720
One remarkably useful property of a Knowm™ synapse, which renders such a device very appropriate for an adaptive neural network, is that the frequency or magnitude of the electric field determines the connection strength. Thus, the connections that become frequently “activated” become stronger. The question of frequency dependence on synapse formation is actually a question of frequency dependence on alignment and connection formation, and can be viewed from at least two different perspectives.
Before further discussion, it will be helpful to make clear some terminology that will aid in the descriptions of the device. In the following descriptions, an adaptive network is built from one base neuron circuit. Each neuron circuit is fundamentally the same, and the network is build by connecting the base neuron circuits together to form certain topologies that result in desired properties, such as maximizing internal feed-back and memory retention. Thus, a complete description can almost be made by describing in detail the function and circuitry of an individual neuron circuit, and then studying how large numbers of the same base neuron will interact with each other. When the term “Neuron” is used, it refers to the electrical analog of a biological neuron, not a biological neuron. This includes summation properties, in time and space, and output properties (e.g., the ability to generate a relatively low-impedance output signal).
Biologically, each neuron can receive signals from other neurons at its post-synaptic terminals. Likewise, in a Knowm™ physical neural network, each neuron can receive signals form it post-synaptic electrodes. Biologically, a neuron transfers signals to other neurons via its pre-synaptic terminals. Likewise, a Knowm™ physical neural network transfers signals via its pre-synaptic electrodes.
The connections between the pre-synaptic electrodes of one neuron and the post-synaptic electrodes of another neuron are formed via nano-connections, and these connections can be seen as independent from either neuron. In other words, the connections do not belong to either neuron, but aid in the transfer of signals form one neuron to another. For example, when one says “positive activation of a pre-synaptic electrode”, one is simply saying that the pre-synaptic electrode is raised to a positive voltage. With these clarifications, we can now proceed.
An adaptive network based on nanotechnology fabrication techniques can be based on, for example, the use of gold nanowires. Gold nanowires are not considered a limiting feature of the present invention, but are described herein for general illustrative and edification purposes only and also to indicate one possible embodiment of the present invention. Gold particles of gold ranging in a diameter of approximately 15 nm to 30 nm can be placed between electrodes deposited on a surface. When an alternating voltage is applied to the electrodes, thin metallic fibers begin to grow on the electrode edge facing the gap.
The fibers can grow in the direction of the other electrode until the gap is bridged, with the wires remaining in contact after the electric field is removed. Nanowire growth can be caused by particle aggregation at the tip of the fibers, thereby extending them toward the opposite electrode. The tip of the growing nanowire can create local electric fields of high intensity and gradient, giving rise to a dielectrophoretic force, which causes the aggregation.
Thus, a first perspective of synapse connection can be seen as a bridge-building process, occurring from one electrode to the other. This process can also be implemented utilizing carbon nanoparticles, such as carbon nanotubes and/or carbon nanowires. Because of their exceedingly small size, carbon nanotubes present a promising possibility because nanotubes have been found to form connections between electrodes. As the frequency increases from 0 Hz to 10's or even 100's of Megahertz, the standard deviation of angles of nanotubes from the electric field decreases. Thus, instead of a bridge-building process, universal alignment of all nanotubes can be implemented.
The nanotubes and/or nanowires and/or other molecular conductors (e.g., molecules), however, may join end to end to bridge a gap, because many carbon nanotubes may overlap between two or more electrodes to form long ropes. With this in mind, the space between two electrodes can be viewed as a multiplicity of bridges between electrodes (conduits), each separated by a characteristic distance that is a result of the local disturbed electric field around each “bridge”. Likewise, many conduits can bundle together to form a rope, bridging the electrodes.
By configuring nanotechnology-based neural circuitry, as described herein, which activates the connections required to be strengthened more frequently, while leaving the connections that need to be weakened inactivated, an adaptive network can be directly emulated utilizing artificial synapses that can both compute and store weight values. All that is required for such a network is a special type of neuron, modeled very much like that of a biological neuron. Each neuron can contain a number of separate functions. By connecting enough neurons into a topology that allows internal feedback, a modifiable network can be constructed which learns though an adaptive feedback process called adaptive integration learning.
As will be discussed later, these neurons can be constructed in conjunction with Knowm synapses to form highly interconnected networks, which use very little space on a VLSI chip. As one will see, such synapses lead themselves to a vertical stacking of planar chips, creating very high-density neural networks. The primary feature of an adaptive network is that of Hebbian learning. If the pre-synaptic neuron fires in close temporal proximity to the post-synaptic neuron, then the connections between the two (i.e., the synapse) can be strengthened. Similarly, if a synapse remains inactive for a long duration, then that connection can be gradually weakened.
Although this mechanism is by no means proven biologically, it provides a possible mechanism of how a cell can strengthen connections in a Hebbian manner, and can be considered the basis for an artificial neuron, which in turn can be used to build an adaptive network. In any case, some mechanism must be in place within a neuron so that its post-synaptic junctions know when they have contributed to neural activation. In an adaptive network utilizing Knowm™ synapses, for example, neurons that achieve a higher rate of activation can produce a higher-frequency (and magnitude) electrical field across their synaptic connections, thereby strengthening such connections. As one can see, the property of Knowm™ synapse strengthening in proportion to an increase in the frequency of an electric field is seminal to the incorporation of a Knowm™ device or component into an adaptive network.
Although the connection modification process is the result of an overall applied frequency, it can also be seen as a very small incremental change for every activation of the synapse. An activation of a synapse can be seen as an activation of a pre-synaptic neuron at the same time as the activation of the post-synaptic neuron. This results in an increase in the electric field, and although the connection strengthening process works with applied frequencies, it can be regarded as an incremental change for every activation to aid in understanding the behavior of the circuit since it is very difficult to picture anything more than small time intervals when dealing with large networks operating at high frequencies and summing signals in a temporal manner over thousands of synaptic inputs.
Signals comprising a voltage from a connection network can be summed by a neuron, in a spatial and temporal manner, and compared to the threshold voltage. One should note that although the voltage from the connection network is shown forming a voltage divider with Rb, which then in turn is summed by the neurons summing circuits, any circuit that accomplishes the same take may be used. For example, an amplifying stage may be added, or the integrating function may become part of the temporal summing circuit. The components described herein are meant to outline basic parts for circuit operation, but are not intended to limit the scope and type of circuit embodiments and implementations thereof.
Evaluating an individual adaptive neuron, if the current state of activation exceeds the threshold voltage, then a pulse may be emitted. During this pulse, the RPG, “refractory pulse generator” grounds the post-synaptic electrodes. This causes the synapses receiving pre-synaptic activation to experience an increase in the local electric field. For example, suppose a pre-synaptic neuron just fired, and caused the firing of the neuron. This means that the pre-synaptic electrode, which itself is connected to the pre-synaptic neuron which just fired, is now on the positive swing of the output pulse.
The neuron, once fired, can output the same pulse, but the RPG will turn it into a negative-going pulse at its post-synaptic electrode. Because the firing of the pre-synaptic neuron precedes that of the post-synaptic neuron, the respective synapse can see (i.e., experience) an alternating electric field. Thus, synapses that contribute to the activation of the neuron can receive both an increase in the local electric field parallel to the connection direction and, when applied many times over, can also experience a higher frequency of activation, two parameters that increase the strength of a Knowm™ synapse. For reasons that will become clear later, the neuron is allowed to source current only on the positive portion of the pulse and sink current on the negative portion of the pulse, where as the neuron cannot source or sink current if it is not activated (i.e., producing a pulse).
One important point to understand before we continue is that each neuron operates in a completely asynchronous mode. This results in every neuron being completely independent from the rest of the network, and consequently massively parallel networks can be built that rely on the emergent behavior of all the interconnected, independent neurons. General properties of the neurons in the network, such as threshold, refractory period and habituation, may be controlled externally via an external CPU. Such external inputs may affect how the network computes, but are not a source of computation in and of themselves.
Before a neuron can act on information coming from its post-synaptic activations, the signals must all be summed. This summation can be performed with, for example, a resistor, Rb, acting as a voltage divider. Summation can alternatively be accomplished with an operational amplifier circuit, which has the added benefit that parameters can be manipulated remotely. A summation circuit can lead to the ability to easily form inhibitory connections and even control the activity of excitatory and inhibitory connections by adjusting the gain of the excitatory and/or inhibitory amplifiers. In a physical chip structure, the area taken (on the chip) for the implementation of a summation circuit, or the use of a one or more large resistor such as Rb, can be a deciding factor in what type of circuit is utilized. Many summations circuits exist, and it is anticipated that a circuit that offers both external control and low component count will be most desirable.
One of the most important features of an adaptive Knowm™ neuron is an integrator. The integrator can sum the signals in time, so that a signal received from one synapse in one instant can be added to a signal from another synapse a short time later. The integrator one uses in an adaptive neuron has a large influence on the behavior of the network. A good analogy is to consider a barrel with a small hole in the bottom, and a trigger that opens a large valve at the bottom of the barrel when the level of incoming water reaches a certain point or threshold. Thus, we can picture pulses of water filling the barrel, from various sources, and a constant leakage of water out of the barrel due to the small hole. If the rate of water into the barrel is greater than the leakage due to the hole, the water level will rise until it hits a point where the valve is triggered and the water is rapidly flushed from the barrel.
In this analogy, the integrator can be seen as the barrel (which stores inputs from past time periods) and the leakage hole, which serves to keep the integrator from accumulating water and firing over long lengths of time when little activation is present. Both of these parameters can be adjusted. The integrator can be built in a number of different ways, and it is the intent of this patent to cover all possible cases, with no preference toward any particular circuit. It is anticipated, for example, that the capacitance of the post-synaptic electrodes could be used as a stage in the construction of an integrator.
Some integrating circuits have been found to be more stable than others. The exact details of the integration circuit are not important for the description of the device described herein, in accordance with one possible embodiment of the present invention, because it is the intent of such an embodiment to cover generally all possible integrator circuits. It is believed that the type of integrator utilized can have a large effect on the performance of a network. Bi-stable integrators, for example, can result in much more robust integrators than previous models. Thus, a bi-stable property may be emulated in the electronic circuitry of an integrator and incorporated into an adaptive neuron.
The next important sub-system of an adaptive neuron is a threshold circuit. This can be accomplished via a number of ways, but a comparator (i.e. Op-amp) provides the simplest example. If the output from the integrator circuit reaches a voltage equal or greater than a threshold voltage, the threshold circuit outputs a signal, which we will assume to be high but could also be low. The threshold circuit, in combination with the integrator and summation circuit, performs the temporal and spatial summations necessary for an adaptive neuron.
The next sub circuit is the pulse-generator. Because Knowm™ networks respond to applied frequencies of electrical fields, it is necessary to encode outputs via pulses. This idea fits nicely with the biological analogy where signals are transmitted as a series of action potentials traveling through axons and dendrites. The input to the pulse generator is the output from the threshold circuit. The output from the pulse generator is, as can be expected, a pulse. The width of the pulse can be determined by the designer of the circuit, or controlled externally.
For reasons that will become clear in the context of an adaptive network, the output of the RPG can be a high (+Vcc) pulse followed immediately by a low (−Vcc or ground) pulse if the neuron output pulse is a low pulse (−Vcc or ground) followed by a high (+Vcc) pulse. Likewise, the output of the RPG can be a low pulse (−Vcc or ground) followed by a high (+Vcc) pulse if the neuron output pulse is a high (+Vcc) pulse followed immediately by a low (−Vcc or ground) pulse.
The output of the adaptive neuron, as far as any post-synaptic neurons are concerned, is generally (although not necessarily) that of the pulse generator, with one small caveat. To implement the form of Hebbian learning via a refractory pulse, as mentioned previously, the adaptive neuron should be allowed to strengthen those post-synaptic synapses that are activated while the neuron is also activated. In other words, pre-synaptic neuron firings that are highly correlated with post-synaptic neuron firings should be strengthened. To accomplish this, a refractory pulse generator can be introduced. The refractory pulse generator takes a positive input (+Vcc from the pulse generator output) and produces a negative (−Vcc or ground) pulse at the post-synaptic electrodes. Following the negative pulse, the refractory pulse generator creates a positive (+Vcc) pulse, which can serve to positively activate the post-synaptic connections immediately after the negative activation.
Alternately, the refractory pulse generator can produce a positive pulse followed by a negative pulse. If the output of the pulse generator is a positive pulse followed by a negative pulse, then the output of the refractory pulse generator can be a negative pulse followed by a positive pulse. The width of this positive pulse can be adjusted, but it can be assumed to be the same as the negative pulse to aid in the description. One can also think of the refractory pulse generator as an inverter of the pulse generator, the output of which projects to the post-synaptic electrode.
Connected to the pulse generator and refractory pulse generator and neuron output are two important sub-circuits, a Selective Current Sink and a Selective Current Source, which serve an important, although not immediately obvious purpose. When the pulse generator outputs a positive pulse, the Selective Current Source allows the neuron to source current. When the refractory pulse is negative, the selective current sink allows the neuron to sink current. If no pulse is present, i.e., the pulse generator is outputting zero, then the current sink and current source does not sink or source current, but leaves the pre-synaptic electrodes floating.
The importance of the Selective Current Sink will become clear when one considers the group behavior of many neurons in a perpendicular array structure, all highly interconnected. Such restrictions on adaptive neurons can restrict current flow to predominantly the pre-to-post synaptic electrode direction, which can also keep unwanted current flows and voltage drops occurring from activated post-synaptic electrodes to inactivated post-synaptic electrodes and also from activated pre-synaptic electrodes to inactivated pre-synaptic electrodes.
In biological neural networks, habituation of the individual neurons plays a large role in global network function. As such, it can be appreciated that such an electrical analog could provide useful computational properties. Biologically, a neuron needs to consume chemical resources to provide the energy to fire. When the resources run low, and the by-products overwhelm the cell, the neurons firing rate begins to slow down. Likewise, when the neuron has not fired for awhile, the chemical resources needed for energy production begin to stockpile, which causes them to fire at heightened frequencies in comparison to neurons that have fired more frequently and not built up chemical reserves. The electrical analog could be provided in many ways, such as making the threshold of the neuron a function of the neurons past firing history. This can be accomplished with digital and/or analog circuitry, as long as the synaptic electrodes receive the proper pulse.
On a superficial level, it ca be appreciated how such a Hebbian learning circuit generally functions. When pre-synaptic neurons activate the adaptive neuron, the refractory pulse generator grounds (i.e., or lowers to −Vcc) all of the post-synaptic electrodes, which can cause an increased electrical field across all connections with high pre-synaptic electrodes. Immediately after the negative activation, the refractory pulse generator positively activates the post-synaptic electrodes. Those pre-synaptic neurons just activated can sink a current for a brief time, because of the selective current sink at their output. Thus, the connections participating in the activation of the neuron will see a full-wave alternating electrical field of increased magnitude. When this process is repeated, selective connections (i.e., those with a temporal correlation in firing) will generally experience an increase in the strength and frequency of the electric filed, and consequently become stronger.
Synaptic connections that activate just before a neuron's activation becomes stronger can experience an increased alternating electric field parallel to connection direction. Additionally, synapses that fire after the neuron activates become weaker (via columbic repulsion), and connections that could result in more efficient signal transduction synapses that fire and do not activate the neuron become slightly stronger (i.e., experience a half-magnitude alternating electrical field). This last form of connection modification provides a form of connection exploration within the circuit. Without a form of non-Hebbian connection formation, potentially useful connections would never form. In other words, for Hebbian learning to take place, the connections should already exist. Hebbian learning only “picks out” those connections that turn out to be useful, and destroys those that cause undesired outputs.
The pulse emitted from the neuron can also take on a variety of other forms, such as, for example, a sinusoidal pulse, triangular pulse, etc. A general concept of how the electric fields at a synapse functions can be obtained if one assumes that the frequencies of pre- and post-synaptic activations are not the same, and considers the beat-frequencies of the input and output wave-forms present at the input and output electrodes. In this case, the gradient of the voltages at the pre- and post-synaptic electrodes can be approximately equivalent to the electric field. Although the connections themselves play an important role in the local electric field, it can be assumed that the cross-sectional area of the pre- and post-synaptic electrodes is large compared to the size of nano-connections, so that this problem is minimized.
As indicated in the background section of this disclosure, researchers, researchers in the neuro-biological fields have been challenged with a need to develop a computationally efficient algorithm that can emulate a biologically realistic neural network. Specifically, researchers have attempted to develop a method, which would allow the efficient calculation of Spike-Timing Dependent-Plasticity (STDP), while also permitting fully interconnected networks. In STDP, timing between pre- and post-synaptic events can cause a net potentiation (LTP) or a net depression (LTD) of synapses.
Many biological studies have found a relationship between “T”, the inter-spike interval and Δg/g, the fractional change in conductance in the synapses, which would indicate LTP or LTD.
To achieve a computationally efficient STDP algorithm, every neuron outputs a characteristic pulse via its axon, after attaining a threshold value. Simultaneously, the neuron can emit a characteristic pulse via its dendrites.
An update to the synapse can be a function of the pre- and post-synaptic pulses. In general, the following rule can be applied: ωt+1=ωt−(pre)(post), where (pre)(post) represents the product of the pre- and post-synaptic activation, as provided by a pulse. Consider the pulses 3100 depicted in
Also, consider the configuration of
A similar but varying format can be followed in
Thus, by varying the pulse shapes, many different learning rules can be implemented. By changing the pulse patterns, post synaptic, pre-synaptic or both, different learning rules can be obtained. For example, by using skewed gaussian pulses in place of square pulses in the previous example, one can obtain an exponential dependence on “T”, instead of a linear. The examples provided herein with respect to the use of pulse generation and pulse shapes are merely examples of embodiments in which the present invention can be implemented are not considered limiting features of the present invention.
An individual nanoconnection may constitute a nanoconductor such as, for example, a nanowire(s), a nanotube(s), nanoparticles(s), or any other molecular structures (e.g., molecules). Nanoconnections 3708 can also be constituted as a plurality of interconnected nanotubes and/or a plurality of interconnected nanowires. Similarly, nanoconnections 3708 can be formed from a plurality of interconnected nanoparticles (i.e. molecules).
A major problem with emulating a biologically realistic neural network models, with pulse-coded outputs and modifiable synapses, concerns the extremely computationally expensive nature of the calculations. Whereas every synapse in a biological network can modify itself individually according to simple rules such as STDP, any calculation must take into account the state of the entire system, in essence storing the state in memory and constantly updating it. Because of the largely serial nature of both modern computers and memory, this task can be extremely inefficient, resulting in simulations that take many hours, days, or even weeks to achieve only a few seconds of simulated “real time”. In addition to the difficulties involved in simulation activities, knowledge of the general structure of biologically realistic networks is also limited.
Although synaptic plasticity rules may create a network that can learn and adapt to its environment, a major step in creating a truly useful network is determining a general structure that allows plasticity to further “tune” the circuit. Such a general structure is, for example, coded in the genome of every living organism. The more simple the organism, the more “hard-wired” the network. It is anticipated that many useful networks will be developed for relatively simple tasks that are currently being accomplished with lower-level organisms such as insects, reptiles and birds.
Evolution has determined general connection patterns over hundreds of millions of years, and encoded this in DNA. Unfortunately, it is not a simple matter of analyzing DNA to determining a general connection pattern. Nor can a brain be efficiently dissected to determine the connection patterns, simply because the fundamental synaptic plasticity rules may differ slightly, and with a different learning rule comes different connection patterns. In designing a pulsed network with modifiable connections, it is very important to allow every possible connection, but it is not desirable that every connection be “on”. For example, the human brain possesses about 100 billion neurons, but each neuron is connected to only 10,000 others.
Knowing which connections should not be connected is a problem that only evolution can solve. This would require the initialization of certain synapses (i.e., turning some synapse on, while leaving the majority of the others off) and then testing the network by subjecting it to stimulus characteristic to the environment in which it will be used. If the network performs the job well, those particular connections can be recorded, and then production may begin only for those connections that are allowed to form. It is very likely, however, that the synapses that are initialized on the first trial will not be adequate. If not, some connections must be turned “on,” “off,” or some combination thereof, perhaps in a random manner that mimics evolution, or perhaps in a systematic manner.
The new pattern of connections must be subjected to its environment, and evaluated as to its ability to perform the desired task. It can be appreciated that a standard computer processor cannot achieve the speed necessary to simulate every generation, because many generations will most likely need to be evaluated, and every generation must have a chance to adapt to its environment via STDP rules. The time required to evaluate one generation via standard computation methods would most likely take many weeks.
In essence, it is necessary to recreate millions of years of evolution and many thousands of generations. To accomplish this, a faster version of the environment should be provided, and the network must adapt at an appropriately speed-up rate. One can accomplish this feat with a Knowm™ adaptive network, but synapses musts be turned “on” and “off” for initialization at the beginning of every generation.
As indicated earlier, computer simulations of neural networks are too complicated and slow for efficient and useful applications. Such serial based algorithms must initialize connections and then run learning routines and then re-initialize, run again, and so forth. Certain connections are desired for initialization, while others simply must be turned off. System 3700 therefore presents a solution to this problem. A gate 3706 can be located adjacent to the connection gap 3710, electrically insulated, and thus provide an electric field for nanoconnections 3708, which can be, for example, semi-conducting nanotubes, nanowires, nanoparticle and/or other molecular semi-conducting structures.
Gate 3706 can be formed from materials, such as, for example, aluminum, gold, and the like. Note that the gate 3706 can be insulated from the nanoconnections and/or other molecular semi-conducting structures with the connection gap 3710 by a material, such as, for example, silicon dioxide or other types of oxide insulators. Although silicon dioxide is shown utilized in the configurations of systems 3700 and 3800, it can be appreciated that other types of insulating materials may also be utilized in place of silicon dioxide and that the use of silicon dioxide is not a limiting feature of the embodiments disclosed herein. Other types of insulating material that could be utilized in place of silicon dioxide or which could potentially complement the use of silicon dioxide as an insulator include materials such as, for example, aluminum oxide, hafnium oxide, zirconium oxide, yttrium oxide and/or silicon dioxide mixed with transition or rare-earth metals such as zirconium and/or lanthanum. The use of silicon dioxide as an insulating material in
Thus, system 3700 is comprised of semi-conducting nanoconnections 3708, rather than simply conducting nanoconnections and/or other conducting molecular structures. The semi-conducting nanoconnections 3708 can be formed from material such as, for example, carbon, silicon, indium phosphide, and so forth. An AC field can be formed across the connection gap 3710, and made to vary, thereby strengthening or weakening nanoconnections and accomplishing STDP rules.
Such a structure can be viewed as analogous to a field-effect transistor, in which the source and drain electrodes correspond to the pre- and post-synaptic electrodes, and which the gate effectively controls if the connection is on or off. What is different about this structure when compared to prior literature is that the resistance between the source and drain can be modified by aligning the nanoparticles with an alternating electric field across the source and drain electrodes, thereby implementing the functionality of a synapse with the added ability that the synapse can be turned “on” or “off” with the gate voltage.
Logic circuitry 3718 can thus include additionally circuitry such as transistors, resistors, capacitors, and the like. The use of gate 3706 in association with nanoconnections 3708 permits the connections 3708 and/or connection network thereof to function as a transistor or a group of transistors for determining which individual synapses or groups thereof are to be identified as activated or deactivated. System 3700 of
System 3900 generally includes a physical neuron 3902 which may form part of a physical neural network as indicated herein. Neuron 3902 can be formed utilizing nanotechnology and/or integrated circuit fabrication techniques as indicated herein. A plurality of weighted factors 3906 (i.e., “w”) can be input to neuron 3902. In general, an input 3908 (i.e., “x”), which is also equivalent to pre-synaptic activity (e.g., pre-synaptic electrode activity) can be multiplied and summed to produce a value, y, which represents neuron activity, as indicated at block 3910. A delta weight factor Δw can be calculated via a Hebbian learning rule, which is indicated at block 3912. An anti-Hebbian learning rule is indicated at block 3914.
Ultimately, an update function f(y), which is indicated at block 3916, can be multiplied by xy and a learning factor to produce a value Δw which is updated based on the function f(y). Based on system 3900, it can be appreciated that synaptic learning amounts to an unsupervised rule that changes synaptic weight as a function of pre- and post-synaptic activity. Function f(y), for example, can be viewed as a mechanism that implements either Hebbian or anti-Hebbian learning based on post-synaptic neural activity. In terms of a physical neural network as described herein, pre/post synaptic activity can be a voltage, frequency or a combination thereof. Thus, a voltage gradient dependency can be utilized to implement Hebbian/Anti-Hebbian learning in a physical neural network based on nanotechnology as indicated herein. Alternatively, varying of pre/post synaptic frequency can also provide for Hebbian/Anti-Hebbian learning.
Based on the foregoing, it can be appreciate that embodiments disclosed herein describe a physical neural network that can be configured utilizing nanotechnology. Such a physical neural network can comprise a plurality of molecular conductors (e.g., nanoconductors) which form neural connections between pre-synaptic and post-synaptic components of the physical neural network. Additionally, a learning mechanism can be applied for implementing Hebbian learning via the physical neural network. Such a learning mechanism can utilize a voltage gradient or voltage gradient dependencies to implement Hebbian and/or anti-Hebbian plasticity within the physical neural network. The learning mechanism can also utilize pre-synaptic and post-synaptic frequencies to provide Hebbian and/or anti-Hebbian learning within the physical neural network.
Note that all of the aforementioned references indicated herein are referred to for general illustrative and edification purposes only and do not disclose or teach the novel invention(s) disclosed and claimed herein.
The embodiments and examples set forth herein are presented to best explain the present invention and its practical application and to thereby enable those skilled in the art to make and utilize the invention. Those skilled in the art, however, will recognize that the foregoing description and examples have been presented for the purpose of illustration and example only. Other variations and modifications of the present invention will be apparent to those of skill in the art, and it is the intent of the appended claims that such variations and modifications be covered.
The description as set forth is not intended to be exhaustive or to limit the scope of the invention. Many modifications and variations are possible in light of the above teaching without departing from the scope of the following claims. It is contemplated that the use of the present invention can involve components having different characteristics. It is intended that the scope of the present invention be defined by the claims appended hereto, giving full cognizance to equivalents in all respects.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US2707223||Jun 15, 1949||Apr 26, 1955||Hans E Hollmann||Electric resistor|
|US3833894||Jun 20, 1973||Sep 3, 1974||Ibm||Organic memory device|
|US4802951||Dec 23, 1987||Feb 7, 1989||Trustees Of Boston University||Method for parallel fabrication of nanometer scale multi-device structures|
|US4926064||Jul 26, 1989||May 15, 1990||Syntonic Systems Inc.||Sleep refreshed memory for neural network|
|US4974146||May 6, 1988||Nov 27, 1990||Science Applications International Corporation||Array processor|
|US4988891||Sep 13, 1989||Jan 29, 1991||Mitsubishi Denki Kabushiki Kaisha||Semiconductor neural network including photosensitive coupling elements|
|US5315162 *||Oct 4, 1991||May 24, 1994||Hughes Aircraft Company||Electrochemical synapses for artificial neural networks|
|US5422983||Jul 19, 1993||Jun 6, 1995||Hughes Aircraft Company||Neural engine for emulating a neural network|
|US5475794||Aug 17, 1994||Dec 12, 1995||Mitsubishi Denki Kabushiki Kaisha||Semiconductor neural network and operating method thereof|
|US5589692||Apr 11, 1995||Dec 31, 1996||Yale University||Sub-nanoscale electronic systems and devices|
|US5649063||Dec 16, 1994||Jul 15, 1997||Lucent Technologies Inc.||Feedback process control using a neural network parameter estimator|
|US5670818||Aug 16, 1994||Sep 23, 1997||Actel Corporation||Electrically programmable antifuse|
|US5706404||Jul 30, 1996||Jan 6, 1998||U.S. Philips Corporation||Neural network using inhomogeneities in a medium as neurons and transmitting input signals in an unchannelled wave pattern through the medium|
|US5717832||Jun 7, 1995||Feb 10, 1998||International Business Machines Corporation||Neural semiconductor chip and neural networks incorporated therein|
|US5761115||May 30, 1996||Jun 2, 1998||Axon Technologies Corporation||Programmable metallization cell structure and method of making same|
|US5783840||Jun 7, 1995||Jul 21, 1998||Texas Instruments Incorporated||Universal quantum dot logic cell|
|US5812993||Feb 27, 1997||Sep 22, 1998||Technion Research And Development Foundation Ltd.||Digital hardware architecture for realizing neural network|
|US5896312||Jan 7, 1998||Apr 20, 1999||Axon Technologies Corporation||Programmable metallization cell structure and method of making same|
|US5904545||Jun 7, 1995||May 18, 1999||The Regents Of The University Of California||Apparatus for fabricating self-assembling microstructures|
|US5914893||Jan 7, 1998||Jun 22, 1999||Axon Technologies Corporation||Programmable metallization cell structure and method of making same|
|US5951881||Jul 22, 1996||Sep 14, 1999||President And Fellows Of Harvard College||Fabrication of small-scale cylindrical articles|
|US5978782||Jul 5, 1996||Nov 2, 1999||National Semiconductor Corporation||Neural network signal processor for magnetic storage channels|
|US6026358||Dec 21, 1995||Feb 15, 2000||Justsystem Corporation||Neural network, a method of learning of a neural network and phoneme recognition apparatus utilizing a neural network|
|US6084796||Jan 12, 1999||Jul 4, 2000||Axon Technologies Corporation||Programmable metallization cell structure and method of making same|
|US6128214||Mar 29, 1999||Oct 3, 2000||Hewlett-Packard||Molecular wire crossbar memory|
|US6245630||May 29, 1998||Jun 12, 2001||Ball Semiconductor, Inc.||Spherical shaped semiconductor integrated circuit|
|US6248529||May 20, 1999||Jun 19, 2001||Integrated Nano-Technologies, Llc||Method of chemically assembling nano-scale devices|
|US6256767||Mar 29, 1999||Jul 3, 2001||Hewlett-Packard Company||Demultiplexer for a molecular wire crossbar network (MWCN DEMUX)|
|US6282530||Jun 9, 1999||Aug 28, 2001||Helios Semiconductor Inc.||Digital neural node|
|US6294450||Mar 1, 2000||Sep 25, 2001||Hewlett-Packard Company||Nanoscale patterning for the formation of extensive wires|
|US6314019||Mar 29, 1999||Nov 6, 2001||Hewlett-Packard Company||Molecular-wire crossbar interconnect (MWCI) for signal routing and communications|
|US6330553||Apr 9, 1998||Dec 11, 2001||Yamaha Hatsudoki Kabushiki Kaisha||Autonomic system for updating fuzzy neural network and control system using the fuzzy neural network|
|US6335291||Nov 24, 1999||Jan 1, 2002||Ball Semiconductor, Inc.||System and method for plasma etch on a spherical shaped device|
|US6339227||Feb 1, 1999||Jan 15, 2002||The Mitre Corporation||Monomolecular electronic device|
|US6359288||Apr 22, 1998||Mar 19, 2002||Massachusetts Institute Of Technology||Nanowire arrays|
|US6363369 *||Jun 11, 1998||Mar 26, 2002||University Of Southern California||Dynamic synapse for signal processing in neural networks|
|US6383923||Aug 22, 2000||May 7, 2002||Agere Systems Guardian Corp.||Article comprising vertically nano-interconnected circuit devices and method for making the same|
|US6389404||Dec 30, 1998||May 14, 2002||Irvine Sensors Corporation||Neural processing module with input architectures that make maximal use of a weighted synapse array|
|US6407443||Jun 20, 2001||Jun 18, 2002||Hewlett-Packard Company||Nanoscale patterning for the formation of extensive wires|
|US6418423||Jan 29, 1999||Jul 9, 2002||International Business Machines Corporation||Method and apparatus for executing neural network applications on a network of embedded devices|
|US6420092||Oct 13, 1999||Jul 16, 2002||Cheng-Jer Yang||Low dielectric constant nanotube|
|US6422450||Sep 15, 2000||Jul 23, 2002||University Of North Carolina, The Chapel||Nanotube-based high energy material and method|
|US6423583||Jan 3, 2001||Jul 23, 2002||International Business Machines Corporation||Methodology for electrically induced selective breakdown of nanotubes|
|US6424961||Dec 6, 1999||Jul 23, 2002||AYALA FRANCISCO JOSť||Adaptive neural learning system|
|US6426134||Jun 29, 1999||Jul 30, 2002||E. I. Du Pont De Nemours And Company||Single-wall carbon nanotube-polymer composites|
|US6445006 *||Jul 27, 1999||Sep 3, 2002||Advanced Technology Materials, Inc.||Microelectronic and microelectromechanical devices comprising carbon nanotube components, and methods of making same|
|US6536106 *||Jun 30, 2000||Mar 25, 2003||The Penn State Research Foundation||Electric field assisted assembly process|
|US6620346||Aug 11, 1998||Sep 16, 2003||Hydro-Quebec||Varistors based on nanocrystalline powders produced by mechanical grinding|
|US6687686 *||Jul 6, 2000||Feb 3, 2004||Massachusetts Institute Of Technology||Hebbian synapse circuit|
|US6763340 *||Sep 22, 2000||Jul 13, 2004||The United States Of America As Represented By The Secretary Of The Air Force||Microelectromechanical system artificial neural network device|
|US6798692||May 21, 2002||Sep 28, 2004||Axon Technologies Corporation||Programmable sub-surface aggregating metallization structure and method of making same|
|US6855329||Jan 20, 1999||Feb 15, 2005||Massachusetts Institute Of Technology||Surface coating spatially controlled patterns|
|US20010004471||Dec 18, 2000||Jun 21, 2001||Nec Corporation||Method of processing a nanotube|
|US20010023986||Feb 7, 2001||Sep 27, 2001||Vladimir Mancevski||System and method for fabricating logic devices comprising carbon nanotube transistors|
|US20010024633||Mar 15, 2001||Sep 27, 2001||Young-Hee Lee||Method of vertically aligning carbon nanotubes on substrates at low pressure and low pressure using thermal chemical vapor deposition with DC bias|
|US20010031900||Mar 16, 2001||Oct 18, 2001||Margrave John L.||Chemical derivatization of single-wall carbon nanotubes to facilitate solvation thereof; and use of derivatized nanotubes to form catalyst-containing seed materials for use in making carbon fibers|
|US20010041160||Mar 16, 2001||Nov 15, 2001||Margrave John L.||Chemical derivatization of single-wall carbon nanotubes to facilitate solvation thereof; and use of derivatized nanotubes to form catalyst-containing seed materials for use in making carbon fibers|
|US20010044114||May 17, 2001||Nov 22, 2001||Integrated Nano-Technologies, Llc.||Chemically assembled nano-scale circuit elements|
|US20020001905||Jun 27, 2001||Jan 3, 2002||Choi Won-Bong||Vertical nano-size transistor using carbon nanotubes and manufacturing method thereof|
|US20020004028||Mar 16, 2001||Jan 10, 2002||Margrave John L.||Chemical derivatization of single-wall carbon nanotubes to facilitate solvation thereof; and use of derivatized nanotubes to form catalyst-containing seed materials for use in making carbon fibers|
|US20020004136||Jun 14, 1999||Jan 10, 2002||Yufei Gao||Carbon nanotubes on a substrate|
|US20020030205||Aug 9, 1999||Mar 14, 2002||Victor I. Varshavsky||Artificial neuron on the base of beta-driven threshold element|
|US20020075126||Oct 3, 2001||Jun 20, 2002||Reitz Hariklia Dris||Multiple reactant nozzles for a flowing reactor|
|US20020086124||Mar 16, 2001||Jul 4, 2002||Margrave John L.|
|US20020090468||Oct 30, 2001||Jul 11, 2002||Honda Giken Kogyo Kabushiki Kaisha||Method of manufacturing carbon nanotube|
|US20020102353||Mar 20, 2002||Aug 1, 2002||Electrovac, Fabrikation Electrotechnischer Spezialartikel Gesellschaft M.B.H.||Method of producing a nanotube layer on a substrate|
|US20030031438||Feb 25, 2002||Feb 13, 2003||Nobuyuki Kambe||Structures incorporating polymer-inorganic particle blends|
|US20030177450||Mar 12, 2002||Sep 18, 2003||Alex Nugent||Physical neural network design incorporating nanotechnology|
|EP1022764A1||Jan 19, 2000||Jul 26, 2000||Lucent Technologies Inc.||Article comprising enhanced nanotube emitter structure and process for fabricating article|
|EP1046613A2||Apr 25, 2000||Oct 25, 2000||Iljin Nanotech Co., Ltd.||Method of forming carbon nanotubes|
|EP1069206A2||Jul 3, 2000||Jan 17, 2001||Lucent Technologies Inc.||Nanoscale conductive connectors and method for making same|
|EP1100106A2||Oct 9, 2000||May 16, 2001||Lucent Technologies Inc.||Article comprising aligned nanowires and process for fabricating article|
|EP1115135A1||Jan 5, 2001||Jul 11, 2001||Samsung SDI Co., Ltd.||Method for fabricating triode-structure carbon nanotube field emitter array|
|EP1134304A2||Mar 15, 2001||Sep 19, 2001||Lee, Young-hee||Method of vertically aligning carbon nanotubes on substrates using thermal chemical vapor deposition with dc bias|
|RU2071126C1||Title not available|
|WO2000044094A1||Jan 20, 2000||Jul 27, 2000||Univ South Carolina||Molecular computer|
|WO2003017282A1||Aug 13, 2001||Feb 27, 2003||Juri Heinrich Krieger||Memory cell|
|1||"A Basic introduction to Neural Networks"; http://blizzard.qis.uiuc.edu/htmldocs/Neural/neural.html.|
|2||"A New Class of Nanostructure: Semiconducting Nanobelts Offer Potential for Nanosensors and Nanoelectronics," Mar. 12, 2001, http://www.sciencedaily.com/releases/2001/03/010309080953.htm.|
|3||"Collective Transport in Arrays of Small Metallic Dots" A. Alan Middleton, N.S. Wingreen, 1993, The American Physical Society, 0031-9007/93/71(19)/3198(4), pp. 3198 through 3201.|
|4||*||"Computational nanotechnology with carbon nanotubes and fullerenes", Srivastava, D. Menon, M. Kyeongjae Cho, NASA Ames Res. Center, Moffett Field, CA, Computing in Science & Engineering, Jul./Aug. 2001, vol. 3, Issue 4, pp. 42-55.|
|5||"Elements of Artificial Neural Networks" K. Mehrotra, C. K. Mohan, S. Ranka, 1997, MIT Press, pp. 116-135.|
|6||"Neural Networks," StatSoft, Inc., http://www.statsoftinc.com/textbook/stevnet.html.|
|7||"Quantum-Dot Arrays for Computation," ORNL Review vol. 34, No. 2, 2001, pp. 1-5.|
|8||"Self-assembled chains of graphitized carbon nanoparticles" A. Bezryadin, R.M. Westervelt, M. Tinkham, Dec. 21, 1998.|
|9||*||"Solid-state thin-film memistor for electronic neural networks", S. Thakoor, A. Moopenn, T. Daud, and A. P. Thakoor, Journal of Applied Physics-Mar. 15, 1990-vol. 67, Issue 6, pp. 3132-3135.|
|10||A. Bezryadin, et al., "Electrostatic trapping of single conducting nanoparticles between nanoelectrodes," Applied Physics Letters, 71 (9), Sep. 1, 1997, pp. 1273-1275.|
|11||A. Bezryadin, Trapping Single Particle with Nanoelectrodes, Physics News Graphics, Sep. 1997.|
|12||A. Jarosz, et al., "An Introductory Note on Gaussian Correlated Random Matrix," Feb. 21, 2003, pp. 1-20 http://www.if.uj.edu.pl/pl/koloSMP/prace/mdmatrix.pdf.|
|13||A. Leonardi, et al., "Simulation methodology for dielectrophoresis in microelectronic Lab-on-a-chip," Modeling and Simulation of Microsystems 2002, pp. 96-99.|
|14||A. van Schaik, "Building blocks for electronic spiking neural networks," Neural Networks 14 (2001), pp. 617-628.|
|15||Abraham Harte, "Liquid Crystals Allow Large-Scale Alignment of Carbon Nanotubes," CURJ (Caltech Undergraduate Research Journal), Nov. 2001, vol. 1, No. 2, pp. 44-49.|
|16||Andriotis et al., "Various bonding configurations of transition-metal atoms on carbon nanotubes: Their effect on contact resistance," Applied Physics Letters, vol. 76, No. 26, Jun. 26, 2000, pp. 3890-3892.|
|17||Avouris et al., "Carbon nanotubes: nanomechanics, manipulation, and electronic devices," Applied Surface Science 141 (1999), pp. 201-209.|
|18||Bezryadin et al., "Evolution of avalanche conducting states in electrorheological liquids," Physical Review E, vol. 50, No. 6, Jun. 1999, pp. 6896-6901.|
|19||Bezryadin et al., "Self-assembled chains of graphitized carbon nanoparticles," Applied Physics Letters, vol. 74, No. 18, May 3, 1999, pp. 2699-2701.|
|20||C.A. Dyke, et al., "Unbundled and Highly Functionalized Carbon Nanotubes from Aqueous Reactions," Nano Letters, 2003, vol. 3, No. 9; Aug. 19, 2003, American Chemical Society, pp. 1215-1218.|
|21||Charles D. Schaper, "Patterned Transfer of Metallic Thin Film Nanostructures by Water-Soluble Polymer Templates," Nano Letters, 2003, vol. 3, No 9; Jul. 26, 2003, American Chemical Society, pp. 1305-1309.|
|22||CMP Cientifica, "Nanotech: the tiny revolution"; CMP Cientifica, Nov. 2001.|
|23||Collins et al., "Engineering Carbon Nanotubes and Nanotube Circuits Using Electrical Breakdown," Science, vol. 292, pp. 706-709, Apr. 27, 2001.|
|24||D.J. Frank, et al., "Highly efficient algorithm for percolative transport studies in two dimensions", Physical Review B, vol. 37, No. 1, Jan. 1, 1988, pp. 302-307.|
|25||Dave Anderson & George McNeill, "Artificial Neural Networks Technology," A DACS (Data & Analysis Center for Software) State-of-the-Art Report, Contract No. F30602-89-C-0082, ELIN: A011, Rome Laboratory RL/C3C, Griffiss Air Force Base, New York, Aug. 20, 1992.|
|26||David Rotman, "Molecular Memory, Replacing silicon with organic molecules could mean tiny supercomputers," Technology Review, May 2001, p. 46.|
|27||David W. Clark, "An Introduction to Neural Networks"; http://members.home.net/neuralnet/introtonn/index.htm.|
|28||Dejan Rakovic, "Hierarchical Neural Networks and Brainwaves: Towards a Theory of Consciousness," Brain & Consciousness: Proc. ECPD Workshop (ECPD, Belgrade, 1997), pp. 189-204.|
|29||Diehl, et al., "Self-Assembled, Deterministic Carbon Nanotube Wiring Networks," Angew. Chem. Int. Ed. 2002, 41, No. 2; Received Oct. 22, 2001.|
|30||Duan et al., "Indium phosphide nanowires as building blocks for nanoscale electronic and optoelectronic devices," Nature, vol. 409, Jan. 4, 2001, pp. 67-69.|
|31||E.S. Snow, et al., "Random networks of carbon nanotubes as electronic material", Applied Physics Letters, vol. 82, No. 12, Mar. 31, 2003, pp. 2145-2147.|
|32||Espejo, et al., "A 16 x 16 Cellular Neural Network Chip for Connected Component Detection," Jun. 30, 1999; http://www.imse.cnm.csic.es/Chipcat/espejo/chip-2.pdf.|
|33||G. Pirio, et al., "Fabrication and electrical characteristics of carbon nanotube field emission microcathodes with an integrated gate electrode," Institute of Physics Publishing, Nanotechnology 13 (2002), pp. 1-4, Oct. 2, 2001.|
|34||Greg Mitchell, "Sub-50 nm Device Fabrication Strategies," Project No. 890-00, Cornell Nanofabrication Facility, Electronics-p. 90-91, National Nanofabrication Users Network.|
|35||Hermanson et al., "Dielectrophoretic Assembly of Electrically Functional Microwires from Nanoparticle Suspensions," Materials Science, vol. 294, No. 5544, Issue of Nov. 2, 2001, pp. 1082-1086.|
|36||Hone et al., "Electrical and thermal transport properties of magnetically aligned single wall carbon nanotube films," Applied Physics Letters, vol. 77, No. 5, Jul. 31, 2000, pp. 666-668.|
|37||J. Appenzeller et al., "Optimized contact configuration for the study of transport phenomena in ropes of single-wall carbon nanotubes," Applied Physics Letters, vol. 78, No. 21, pp. 3313-3315, May 21, 2001.|
|38||J. Chung, et al., "Nanoscale Gap Fabrication and Integration of Carbon Nanotubes by Micromachining," Solid-State Sensor, Actuator and Microsystems Workshop, Jun. 2-6, 2003; Hilton Head Island, South Carolina, pp. 161-164.|
|39||J. Chung, et al., "Nanoscale Gap Fabrication by Carbon Nanotube-Extracted Lithography (CEL)", Nano Letters, 2003, vol. 3, No. 8; Jul. 9, 2003, American Chemical Society, pp. 1029-1031.|
|40||J. Tour, et al., "Nanocell Logic Gates for Molecular Computing," IEEE Transactions on Nanotechnology, vol 1, No. 2, Jun. 2002, pp. 100-109.|
|41||J. Zaumseil, et al., "Three-Dimensional and Multilayer Nanostructures Formed by Nanotransfer Printing," Nano Letters, 2003, vol. 3, No. 9, Jul. 31, 2003, American Chemical Society, pp. 1223-1227.|
|42||J.M. Tour, et al., "NanoCell Electronic Memories," J.Am.Chem.Soc. 2003, 125, pp. 13279-13283.|
|43||Jeong-Mi Moon et al., "High-Yield Purification Process of Singlewalled Carbon Nanotubes," J. Phys. Chem. B 2001, 105, pp. 5677-5681.|
|44||John G. Spooner, "Tiny tubes mean big chip advances," Cnet News.com, Tech News First, Apr. 26, 2001.|
|45||John-William DeClaris, "An Introduction to Neural Networks," http://www.ee.umd.edu/medlab/neural/nn1.html.|
|46||Jouri, M.A. et al., "Adaptive Analog VLSI Neural Systems," Chapman & Hall, London SE1 8HN, UK, 1996, pp. 92-95.|
|47||K. Bradley, et al., "Flexible Nanotube Electronics," Nano Letters, 2003, vol. 3, No. 10; Aug. 9, 2003, America Chemical Society, pp. 1353-1355.|
|48||K. Bradley, et al., "Influence of Mobile Ions on Nanotube Based FET Devices," Nano Letters, 2003, vol. 3, No. 5; American Chemical Society, Apr. 4, 2003; pp. 639-641.|
|49||L. Perrinet, et al., "Coherence detection in a spiking neuron via Hebbian learning," Neurocomputing, 2002, vol. 44-46, No. C., pp. 817-822, http://www.laurent.perrinet.free.fr/publi/perrinet02.pdf.|
|50||L. Perrinet, et al., "Emergence of filters from natural scenes in a sparse spike coding scheme," Neurocomputing, 2003, pp. 1-14, http://www.laurent.perrinet.free.fr/publi/perrinet03neurocomputing.pdf.|
|51||L. Zheng, et al., "Towards Single Molecule Manipulation with Dielectrophoresis Using Nanoelectrodes," IEEE-NANO 2003, Aug. 12-14, 2003, Moscone Convention Center, San Franscisco, CA; pp. 437-440.|
|52||L.A. Nagahara, "Directed placement of suspended carbon nanotubes for nanometer-scale assembly," Applied Physics Letters, vol. 80, No. 20; May 20, 2003; pp. 3826-3828.|
|53||Landman et al., "Metal-Semiconductor Nanocontacts: Silicon Nanowires," Physical Review Letters, vol. 85, No. 9, Aug. 28, 2000.|
|54||Leslie Smith, "An Introduction to Neural Networks," Center for Cognitive and Computational Neuroscience, Dept. of Computing & Mathematics, University of Stirling, Oct. 25, 1996; http//www.cs.stir.ac.uk/~Iss/NNIntro/InvSlides.html.|
|55||Lipson et al., "Automatic Design and Manufacture of Robotic Lifeforms," NATURE, vol. 406, Aug. 31, 2000, pp. 974-978.|
|56||Lu et al., "Fullerene Pipes," Science, vol. 280, May 22, 1998, pp. 1253-1255.|
|57||M. Dubson, et al., "Measurement of the conductivity exponent in two-dimensional percolating networks: square lattice versus random-void continuum", Physical Review B, vol. 32, No. 11, Dec. 1, 1985, pp. 7621-7623.|
|58||M. Law, et al., "Photochemical Sensing of NO<SUB>2 </SUB>with SnO<SUP>2 </SUP>Nanoribbon Nanosensors at Room Temperature," Angew. Chem. 2002, 114, Nr. 13, pp. 2511-2514.|
|59||M. S. Kumar, et al., "Influence of electric field type on the assembly of single walled carbon nanotubes," Chemical Physics Letters 383 (2004), Dec. 2, 2003; pp. 235-239.|
|60||M.C.W. van Rossum, et al., "Stable Hebbian Learning from Spkke-Timing-Dependent Plasticity", The Journal of Neuroscience, Dec. 1, 2003, 20(23), pp. 8812-8821.|
|61||Mark K. Anderson, "Mega Steps Toward the Nanochip," Wired News, Apr. 27, 2001.|
|62||Meyer et al., "Computational neural networks: a general purpose tool for nanotechnology," Abstract, 5<SUP>th </SUP>Foresight Conference on Molecular Nanotechnology; http://www.foresight.org/Conferences/MNT05/Abstracts/Meyeabst.html.|
|63||Nace L. Golding, et al., "Dendritic spikes as a mechanism for cooperative long-term potentiation", NATURE, vol. 418, Jul. 18, 2002, pp. 326-330.|
|64||Niyogi et al., "Chromatographic Purification of Soluble Single-Walled Carbon Nanotubes (s-SWNTs)," J. Am. Chem. Soc 2001, 123, pp. 733-734.|
|65||O. Harnack, et al., "Rectifying Behavior of Electrically Aligned ZnO Nanorods," Nano Letters, 2003, vol. 3, No. 8; Jun. 24, 2003, American Chemical Society, pp. 1097-1101.|
|66||Osamu Fujita, "Statistical estimation of the number of hidden units for feedforward neural networks," Neural Networks 11 (1998), pp. 851-859.|
|67||Ozgur Turel, et al., "Possible nanoelectronic implementation of neuromorphic networks", Dept. o f Physics and Astronomy, Stony Brook University.|
|68||P. O'Connor, G. Gramegna, P. Rehak, F. Corsi, C. Marzocca, CMOS Preamplifier with High Linearity and Ultra Low Noise for X-Ray Spectroscopy, IEEE Transactions on Nuclear Science, vol. 44, No. 3, Jun. 1997, pp. 318-325.|
|69||Pati et al., "Neural Networks for Tactile Perception," Systems Research Center and Dept. of Electrical Engineering, University of Maryland and U.S Naval Research Laboratory. 1987; http://www.isr.umd.edu/TechReports/ISR/1987/TR<SUB>-</SUB>87-123/TR<SUB>-</SUB>87-123.phtml.|
|70||Paulson, et al., "Tunable Resistance of a Carbon Nanotube-Graphite Interface," Science, vol. 290, Dec. 1, 2000, pp. 1742-1744.|
|71||Peter Weiss, "Circuitry in a Nanowire: Novel Growth Method May Transform Chips," Science News Online, vol. 161, No. 6; Feb. 9, 2002.|
|72||Phaedon Avouris, "Molecular Electronics with Carbon Nanotubes," Accounts of Chemical Research, vol. 35, No. 12; Jul. 31, 2002, pp. 1025-1034.|
|73||Press Release, "Nanowire-based electronics and optics comes one step closer," Eureka Alert, American Chemical Society; Feb. 1, 2002.|
|74||Press Release, "Toshiba Demonstrates Operation of Single-Electron Transistor Circuit at Room Temperature," Toshiba, Jan. 10, 2001.|
|75||R. Krupke, "Separation of Metallic from Semiconducting Single-Walled Carbon Nanotubes," Science, vol. 301; Jul. 18, 2003; pp. 344-347.|
|76||R. Krupke, et al., "Contacting single bundles of carbon nanotubes with alternating electric fields," Appl. Phys. A. 76, Oct. 28, 2002, pp. 397-400.|
|77||R. Krupke, et al., "Simultaneous Deposition of Metallic Bundles of Single-walled Carbon Nanotubes Using Ac-dielectrophoresis," Nano Letters, 2003, vol. 3, No. 8; Jul. 9, 2003; American Chemical Society, pp. 1019-1023.|
|78||R. Martel, et al., "Ambipolar Electrical Transport in Semiconducting Single-Wall Carbon Nanotubes," Physical Review Letters, vol. 87, No. 25, Dec. 17, 2001, pp. 256805-1 to 256805-4.|
|79||S. Heinze, et al., "Carbon Nanotubes as Schottky Barrier Transistors", vol. 89, No. 10, Sep. 2, 2002, pp. 106801-1 to 106801-4.|
|80||S. Suzuki, et al., "Quantitative Analysis of DNA Orientation in Stationary AC Electric Fields Using Fluorescence Anisotropy," IEEE Transactions of Industry Applications, vol. 34, No. 1; Jan./Feb. 1998, pp. 75-83.|
|81||S.W. Lee, et al., "Dielectrophoresis and electrohydrodynamics-mediated fluidic assembly of silicon resistors," Applied Physics Letters, vol. 83, No. 18, Nov. 3, 2003, pp. 3833-3835.|
|82||Saito et al., "A 1M Synapse Self-Learning Digital Neural Network Chip," ISSCC, pp. 6.5-1 to 6.5-10, IEEE 1998.|
|83||Smith et al., "Electric-field assisted assembly and alignment of metallic nanowires," Applied Physics Letters, vol. 77, No. 9, Aug. 28, 2000, pp. 1399-1401.|
|84||Smith et al., "Structural anisotropy of magnetically aligned single wall carbon nanotube films," Applied Physics Letters, vol. 77, No. 5, Jul. 31, 2000, pp. 663-665.|
|85||Snow, et al., Nanofabrication with Proximal Probes, Proceedings of the IEEE, Apr. 1997.|
|86||Stephen Jones, "Neural Networks and the Computation Brain or Maters relating to Artificial Intelligence," The Brain Project, http://www.culture.com.au/brain<SUB>-</SUB>proj/neur<SUB>-</SUB>net.htm.|
|87||T.B. Jones, "Frequency-dependent orientation of isolated particle chains," Journal of Electrostatics, 25 (1990), Elsevier Science Publishers, pp. 231-244.|
|88||Tohji et al., "Purifying single walled nanotubes," Nature, vol. 383, Oct. 24, 1996, p. 679.|
|89||Uma R. Karmarkar, et al., "A Model of Spike-Timing Dependent Plasticity: One or Two Coincidence Detectors?", J. Neurophysiol, vol. 88, pp. 507-513, Jul. 2002.|
|90||Uma R. Karmarkar, et al., "Mechanisms and significance of spike-timing dependent plasticity," Biol. Cybern. 87, 373-382 (2002), Jan. 28, 2002.|
|91||V. Derycke et al., "Carbon Nanotube Inter- and Intramolecular Logic Gates," American Chemical Society, Nano Letters, XXXX, vol. 0, No. 0, A-D.|
|92||V.C. Moore, et al., "Individually Suspended Single-Walled Carbon Nanotubes in Various Surfactants," Nano Letters, 2003, vol. 3, No. 10; American Chemical Society; Sep. 8, 2003; pp. 1379-1382.|
|93||V.C. Moore, et al., "Individually Suspended Single-Walled Carbon Nanotubes in Various Surfactants," Nano Letters, 2003, vol. 3; Sep. 9, 2003; American Chemical Society, pp. 1379-1382.|
|94||Weeks et al., "High-pressure nanolithography using low-energy electrons from a scanning tunneling microscope," Institute of Physics Publishing, Nanotechnology 13 (2002), pp. 38-42; Dec. 12, 2001.|
|95||Wei et al., "Reliability and current carrying capacity of carbon nanotubes," Applied Physics Letters, vol. 79, No. 8, Aug. 20, 2001, pp. 1172-1174.|
|96||Westervelt et al., "Molecular Electronics," NSF Functional Nanostructures Grant 9871810, NSF Partnership in Nanotechnology Conference, Jan. 29-30, 2001; http://www.unix.oit.umass.edu/~nano/NewFiles/FN19<SUB>-</SUB>Harvard.pdf.|
|97||Wolfgang Maass, "Noisy Spiking Neurons with Temporal Coding have more Computational Power than Sigmoidal Neurons," In M. Mozer, M. I. Jordan, and T. Petsche, editors, Advances in Neural Information Processing Systems, vol. 9, pp. 211-217. MIT Press (Cambridge), 1997. (pp. 1-13, including Appendix).|
|98||Wolfgang Maass, "On the Relevance of Time in Neural Computation and Learning," In M. Li and A. Maruoka, editors, Proc. of the 8th International Conference on Algorithmic Learning Theory in Sendai (Japan), vol. 1316 of Lecture Notes in Computer Science, pp. 364-388. Springer (Berlin), 1997.|
|99||X. Liu, et al., "Electric-Field Induced Accumulation and Alignment of Carbon Nanotubes," 2002 Annual Report Conference on Electrical Insulation and Dielectric Phenomena, pp. 31-34.|
|100||Xiaohui Xie, et al., "Spike-based learning rules and stabilization of persistent neural activity."|
|101||Yamamoto et al., "Orientation and purification of carbon nantubes using ac electrophoresis," J. Phys. D: Appl. Phys 31 (1998) L34-L36.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7752151 *||Apr 10, 2008||Jul 6, 2010||Knowmtech, Llc||Multilayer training in a physical neural network formed utilizing nanotechnology|
|US7838809 *||Feb 19, 2008||Nov 23, 2010||Ludwig Lester F||Nanoelectronic differential amplifiers and related circuits having carbon nanotubes, graphene nanoribbons, or other related materials|
|US7858918||Feb 4, 2008||Dec 28, 2010||Ludwig Lester F||Molecular transistor circuits compatible with carbon nanotube sensors and transducers|
|US8013286 *||Nov 17, 2010||Sep 6, 2011||Ludwig Lester F||Nanoelectronic differential amplifiers and related circuits having graphene nanoribbons|
|US8112372||Nov 14, 2008||Feb 7, 2012||Christopher D. Fiorello||Prediction by single neurons and networks|
|US8324555||Aug 25, 2011||Dec 4, 2012||Pike Group Llc||Nanoelectronic differential amplifiers and related circuits implemented on a segment of a graphene nanoribbon|
|US8682822||Jun 19, 2012||Mar 25, 2014||International Business Machines Corporation||Electronic learning synapse with spike-timing dependent plasticity using memory-switching elements|
|US8781983||Dec 21, 2010||Jul 15, 2014||Knowmtech, Llc||Framework for the evolution of electronic neural assemblies toward directed goals|
|US8909575||Feb 29, 2012||Dec 9, 2014||Qualcomm Incorporated||Method and apparatus for modeling neural resource based synaptic placticity|
|US8918353||Aug 27, 2012||Dec 23, 2014||Knowmtech, Llc||Methods and systems for feature extraction|
|US8941047||Nov 22, 2010||Jan 27, 2015||Pike Group Llc||Molecular transistor driving of nanoscale actuators from differential amplifier circuits compatible with carbon nanotube sensors and transducers|
|US8972316||Sep 14, 2012||Mar 3, 2015||Knowmtech, Llc||Extensible adaptive classification framework|
|US8983886||Sep 4, 2012||Mar 17, 2015||Knowmtech, Llc||Self-evolvable logic fabric|
|US8990136||Sep 13, 2012||Mar 24, 2015||Knowmtech, Llc||Methods and systems for fractal flow fabric|
|US9053428||Jul 21, 2011||Jun 9, 2015||Qualcomm Incorporated||Method and apparatus of robust neural temporal coding, learning and cell recruitments for memory using oscillation|
|US9064215||Jun 14, 2012||Jun 23, 2015||Qualcomm Incorporated||Learning spike timing precision|
|US9092735||Sep 21, 2011||Jul 28, 2015||Qualcomm Incorporated||Method and apparatus for structural delay plasticity in spiking neural networks|
|US9104975||Mar 15, 2012||Aug 11, 2015||Knowmtech, Llc||Memristor apparatus|
|US9111224||Oct 19, 2011||Aug 18, 2015||Qualcomm Incorporated||Method and apparatus for neural learning of natural multi-spike trains in spiking neural networks|
|US9147155||Aug 16, 2011||Sep 29, 2015||Qualcomm Incorporated||Method and apparatus for neural temporal coding, learning and recognition|
|USRE44469||Apr 11, 2012||Sep 3, 2013||Pike Group Llc||Nanoelectronic differential amplifiers and related circuits having graphene nanoribbons|
|U.S. Classification||706/26, 977/742, 706/6, 706/25, 977/700, 977/712, 706/15|
|International Classification||G06N3/08, G06N3/063, G06E1/00, G06N99/00|
|Cooperative Classification||G06N99/007, Y10S977/70, Y10S977/712, B82Y10/00, Y10S977/742, G06N3/08, G06N3/063|
|European Classification||B82Y10/00, G06N3/08, G06N3/063, G06N99/00M|
|Mar 26, 2012||REMI||Maintenance fee reminder mailed|
|Apr 3, 2012||SULP||Surcharge for late payment|
|Apr 3, 2012||FPAY||Fee payment|
Year of fee payment: 4
|Jan 21, 2013||AS||Assignment|
Owner name: KNOWMTECH, LLC, NEW MEXICO
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:NUGENT, ALEX;REEL/FRAME:029663/0299
Effective date: 20130121